Friday, July 27, 2007

AMD Technology Analysts Day 2007 – quite a surprise.

Since Intel began talking about Penryn we've had a lot of rumors and speculation about K10. With live demonstrations from Intel and no information from AMD, I suppose it is only reasonable to expect rumors to pop up like so many mushrooms. Now that AMD has finally spoken though we can assign most of these to the trashcan. This Technology Analysts Day was different from the one last year. This one was less about initiatives like Torrenza, Trinity, and Raiden and more about actual hardware. It appears we finally have the answer to whether AMD is still going to be around in 2008 or 2009. The answer is clearly, yes.

There are a lot of rumors to take care of but I'll make a couple of points first. For example, lately I've seen people pushing the idea that AMD stuffed the channel in Q4 06 and this then led to the sharp drops in Q1 07. I can't say that I entirely blame these people for such speculation because I've been puzzled about this myself. Every time I tried to understand how AMD got from chip shortages in Q1 to full conversion of FAB 36 to 65nm before the end of Q2, it made no sense. The only explanation was that AMD had pulled off a near miracle and gotten 65nm ramped at FAB 36 in record time. Considering that this would have been faster than Intel converted to 65nm, it was a lot simpler to just assume that AMD was lying. I was a bit surprised therefore to have such speculation actually confirmed by AMD. They stated that they did indeed convert FAB 36 to 65nm in record time and they seemed to give plenty of credit for this to APM.

Another unexpected item was that AMD has pushed Direct Connect 2.0 and Fusion back to 2009. There won't be any processors in 2008 with four HT links or GPU's. And, with only three HT links AMD may still be limited to a best range of 4-way without using glue chips. On the other hand, I had only expected a minor core update in 2008 with DC 2.0. Instead, DC 2.0 will arrive on the brand new Bulldozer core. There isn't much known about it except it has longer pipelines and more powerful SSE performance with additional instructions and better IOMMU for virtualization. From this it appears that AMD is serious about holding onto the HPC top end. Now to the rumors:


Rumor: AMD is selling its FABs and outsourcing CPU production to TSMC.

This piece of FUD started with Covello's note with its wild speculation as a justification for buying more Intel stock. Yet, a number of people on the web have repeated and expanded on these rumors so much that AMD even felt compelled to address it directly:

AMD also explained exactly what "Asset Light" included:

I'm certain that these rumors will now try to morph into the argument that although AMD is not outsourcing CPU's today it probably will in the future. However, AMD was quite clear that FAB 30 would indeed be upgraded to FAB 38 and continue past 32nm:

They were also clear that the only CPU's that will be outsourced are embedded versions of Bobcat which can provide an x86 replacement for embedded MIPS in Xilleon and Imageon.


Rumor: Silverthorne is Intel's silver bullet.

Silverthorne is matched by AMD's Bobcat core in 2009 which draws as little as 1 watt. This now finally explains why AMD unloaded the Geode line so quickly.

Secondly, although Silverthorne should be a good competitor, AMD has twice the Intellectual Property plus relationships already in place with ATI customers for similar products. So Intel will have to work hard to catch up. Of course, Intel's existing relationship with Apple could give it the iPhone market. However iPhone is not likely to become a commodity item according to EETimes.


Rumor: The ATI merger has been a disaster with ATI employees leaving before they get laid off and ATI losing share since the merger.

This rumor was specifically addressed by AMD. It was stated that the attrition rate at ATI has always been low and that it has not increased since the merger. It was further stated that additional engineering staff had to be hired to handle AMD's new chipset and GPU related projects. Finally, the 790 55nm chipset has been pushed forward to Q4 so apparently things are on track.



Rumor: R600 proves that ATI is no longer competitive.

When R600 was reviewed it was on an 80nm process with leaky 80hs transistors. According to AMD, it is now being produced on a 65nm process and this version should be out in Q4. Secondly, few people seem to have noticed the success of the 690G chipset which is being used on 35 different motherboards.



Rumor: AMD's 65nm process is broken and that is why Barcelona can't clock.

At Technology Analysts Day 2007, AMD showed a 3.0Ghz quad core Phenom demo system. The door was open showing that the processor was running with stock air cooling.



Rumor: The SuperPi and Cinnebench tests proved that K10 is really no faster than K8.

In the VMware benchmark, after normalizing for twice as many cores and a slower clock, we see a 34% increase in IPC. This would seem a bit strange because if we combined this with AMD's 3.0Ghz demo we would end up with a system that is not only faster than anything Intel has today but is just as fast as a 3.33Ghz Penryn.

I had to remove the SPECfp graph because AMD has changed this in their official pdf. In the presentation it said that it was an actual test but in the pdf it now says that it is simulated. I had wondered about that when I saw it because it looked similar to the simulated benchmark. However, I just saw an actual comparison between a Barcelona 2.0Ghz system and a Xeon E5345 2.33Ghz system. Barcelona gets 78 on SPECfp while Xeon gets 60. This is 51% greater SSE IPC for K10 than Clovertown.


Rumor: Even if K10 is as fast as Penryn it won't matter because Nehalem's greater speed and massive connectivity will easily beat K10.

Maybe not. K10 works best with 4-way but Shanghai will have 8 cores making for a simple 4-way/32 core system. It also remains to be seen if HT 3.0's greater bandwidth allows reasonable 8-way systems. This would potentially be 64 cores on one system without using any glue chips. Intel loses its MCM advantage since Shanghai also uses MCM. Shanghai only needs half as much cache as Penryn so Intel gains no die size advantage (unless Nehalem has much less cache). Shanghai will also gain in IPC much as Penryn does. There is also some indication that Intel will cancel Tigerton and simply fill in with Tulsa until Nehalem is ready. This seems odd because Caneland will effectively be obsolete (along with all Woodcrest, Clovertown, and Penryn based systems) as soon as Nehalem is released. And, the later Nehalem appears in 2008 the closer it is to Bulldozer.


Rumor: Intel has two teams working on C2D. AMD, with fewer engineers will never be able to keep up with Intel's Tick Tock.

AMD has now announced Pipe which is an identical upgrade cycle to Intel's Tick Tock. It includes the same major core upgrade every two years with process upgrades in the alternating years.



Rumor: AMD doesn't even mention DTX anymore so it must be a failure.

AMD is already counting DTX as a success for 2007:

So, there must be considerable support for DTX. I have to wonder too if DTX (especially mini-DTX) is what AMD has in mind for Fusion. Fusion will allow OEM's to build systems without a northbridge and this which would seem to be a good match for a small form factor.


So, what does all this mean? Apparently AMD's success with the yields at 65nm has allowed them to push back the upgrade schedule for FAB 30. This certainly eases cost pressure on AMD while still allowing them to gain income from 200mm tooling sales. AMD's chipset and CPU lineups also seem competitive. Today is truly nothing like 2003 when AMD introduced Opteron with only its own supporting chipset. K10 is a drop-in replacement for K8 on socket AM2 or socket F and has nearly universal support among OEMs. This should allow K10 to gain traction far more rapidly than the slow pace of 2003 when it took months to even have one desktop chipset announced. It took a full year from its introduction for K8 to surpass K7 but today K10 should surpass K8 in about half that time.

AMD's financial problems are not over, of course, but it should be able to steadily improve its losses over the next three quarters. Although many assume that Intel's price cuts have made it impossible for AMD to make any profit with chip sales, this is not really the case. Intel's prices for C2D have remained relatively high so pricing pressure was more effective when AMD still had the bulk of its production on 90nm. For example, Intel has shown great reluctance to move Conroe prices down into the Celeron range. So, today, Intel is trying to cover the Celeron range with the single core Conroe L 420, 430, and 440 models which are not much of a bargain when matched up with AMD's similarly priced dual core X2 3600, 3800, and 4000 models.

The lowest priced real Conroe is the 1.6Ghz dual core Allendale E2140 for $81. This model is easily surpassed in everything except SSE performance by the 2.2Ghz X2 4200 at $80. Prices for AMD processors remain better than Conroe up to the 2.66Ghz E6700 where its higher speed surpasses AMD's fastest model. Essentially, AMD's X2's are all favorably priced up to and including the 3.0Ghz X2 6000 model at only $170. At current prices, the single Core Conroe L models are not competitive and the dual core models are not as favorably priced as AMD's X2's. Conroe's are a bargain though if your application needs the greater SSE performance of C2D. If you need greater speed then even the fastest 3.0Ghz dual core E6850 at $330 and the lowest priced quad core 2.4Ghz Q6600 at $320 are reasonably priced. Obviously, the prices of the faster quad cores will drop when they face quad core competition from AMD.

I'm certain the 3.0Ghz quad demo left many people wondering when such chips would actually be available. Anandtech's take seems particularly negative suggesting as late as Q2 08. But then Anand hasn't exactly been objective about AMD in the past five years so perhaps we should consider that the upper bound. Realistically, the 3.0Ghz chip could have been cherry picked. And, it generally takes about six months for production to catch up to a cherry picked chip. So, I can't imagine that 3.0Ghz would arrive later than Q1 08. Before that happens though we'll have to find out just how well Barcelona really stacks up to C2D. If the VMWare ratio is genuine then Barcelona would launch with a 2.0Ghz speed equal to a 2.4Ghz Clovertown or a 2.8Ghz Opteron. This would mean a 2.0Ghz dual core Phenom would probably match a slower 2.3Ghz Conroe. These IPC ratios are very important because it will be easier for AMD to produce lower clocked K10's, and if AMD can match C2D at a lower clock then this is good for both AMD's volume and pricing. This will be the central factor in AMD's financial recovery. If K10 needs to match C2D 1:1 in clock to be competitve then AMD will have a very tough time over the next 3 quarters. However, the higher K10's IPC and therefore the less clock speed that K10 needs to match C2D the faster AMD will recover. And, this ratio should be known for certain perhaps in less than one month.

114 comments:

lex said...

Lots of pretty foils with lots of fancy roadmaps that on paper look to answer every INTEL thrust with a response.

A few small matters missing

1) AMD's huge debt that they must service. This requires them to start generating some serious cash

2) The foils look good but they won't translate into the kind of performance advantage AMD enjoyed in late 2005 and early 2006. Thus at the current margins they and MS they won't make enough money to service debt, grow manufacturing capacity AND do the required R&D to enable 45nm and 32nm developement at 2 year cadence. If they pull something togather it will be inferior to INTEL's offering resulting in more pricing pressure for esentially the same fab costs...

Sorry AMD business plan still isn't longterm viable

Aguia said...

According to AMD, it is now being produced on a 65nm process and this version should be out in Q4.

Where did you get that Scientia, I saw the presentation and I also remember that is going to be released some new GPU in Q4 but I’m not sure it was R600.

And, it generally takes about six months for production to catch up to a cherry picked chip.

Well maybe years, maybe never …
Intel Core 2 Extreme demonstrated with 3.5 GHz

Once again, very good article and analysis.

Scientia from AMDZone said...

lex

"AMD's huge debt that they must service. This requires them to start generating some serious cash"

This has been addressed. Did you not understand it?

"Thus at the current margins"

This has been addressed as well.

"they won't make enough money to service debt, grow manufacturing capacity AND do the required R&D to enable 45nm and 32nm developement at 2 year cadence."

AMD is doing 45nm pilot runs right now. But I don't understand your comment about money.

"If they pull something togather it will be inferior to INTEL's offering"

What something are you referring to?

" resulting in more pricing pressure for esentially the same fab costs..."

Resulting in circular reasoning on your part.

"Sorry AMD business plan still isn't longterm viable"

Your comment might be worth something if you put in a few details. What exactly is it that you disagree with?

Scientia from AMDZone said...

aguia

The sample that Intel showed back then was overclocked. And, I've shown that C2D is pushing its thermal limits at 3.0Ghz on either the F2 or F3 stepping. It is possible that the G0 stepping could exceed this.

Given that the AMD demo appears to use stock cooling I assume it is simply a particular good sample rather than running above normal TDP.

Fujiyama said...

Listening AMD management team I'm rather disappointed cause we still hear about upcoming customer centric solutions but we cannot buy them.
Regarding the plan from 2006 Day - and large number of announcements none of the projects was released.

1. Torrenza, Raiden, Trinity - cannot buy them.
2. 65nm Turions X2 - cannot buy them
3. DTX from factor - again
4. HT 3.0 chipsets - ...

So Hector should find a new job cause while the market would like to buy some - AMD is not able to deliver it.

If we look at Barcelona/Agena introduction scheduled for 07- it will take 9 months to launch the whole portfolio for these chips. (June 07-March 08).

AMD is always late in delivery and it doesn't change these days.

Aguia said...

When R600 was reviewed it was on an 80nm process with leaky 80hs transistors. According to AMD, it is now being produced on a 65nm process and this version should be out in Q4.

Scientia you didn’t answer my question, where did you get the 65nm R600, because in the all presentation the only thing we know is one/some GPU(s) that will be released in Q4 at 55nm. Those can be R6xx with 256bit memory controllers and others. Where did you get that info?

Scientia from AMDZone said...

fujiyama

"1. Torrenza, Raiden, Trinity - cannot buy them."

Which of these are not available?

"2. 65nm Turions X2 - cannot buy them"

You've been able to buy these for some time. A 65nm Brisbane 2.6Ghz X2 is $119 at NewEgg.

"3. DTX from factor - again"

This is in the chart: due in Q4.

"4. HT 3.0 chipsets - ..."

Also due in Q4.

Scientia from AMDZone said...

aguia

Slide 148:

Expanded GPU production at 65nm
55nm planned in Q4 2007


The RD790 chipset will be 55nm.

The R7XX series GPU will not be produced until 2008.

However, during the presentation, AMD officials stated more than once that 65nm GPU's were in production now. This is stated again on slide 75:

2007 Successes:
65nm Graphics Processors

It again shows 65nm GPU's on slide 156 along with 55nm at the end of the year.

It is possible that these 65nm GPU's could be available at the end of this quarter but let's say by early Q4.

Azmount Aryl said...

2. 65nm Turions X2 - cannot buy them

No, why, you can buy them, they'll just cost you a little bit extra: LINK

Fujiyama said...

Scientia,
Are you joking?

Where you can buy Torrenza solutions presented in slides last year (i.e. XML, Java, Database). Same for another two 'buzzwords'.

You telling me about Brisbane, I'm talking about Turion X2 CPU for laptops where only 90nm are available.
So find TL-66 CPU in Dell, Toshiba or HP offerings.
According to AMD, these chips are available from March.

InTheKnow said...

Scientia, even though AMD has "addressed" asset lite if find myself less than impressed.

For example it includes "utilization of existing assets to their fullest extent to maximize ROI." What a load of hogwash. Am I supposed to assume that up to this point they did not bother to utilize their assets fully or look at the ROI on every major business decision?

By comparison, when Intel restructured a while back, they put numbers on how much their efforts were going to save each year. That amounted to $2bil for the current year, and they reported they were on track to hit that this year in their latest analyst meeting. Where are the numbers for AMD? Tossing empty market speak around doesn't convince me.

And I don't see anything in what you have posted that says whether or not they are going to outsource CPUs or not. I'll concede it may have been in another slide, but I don't see it here.

Finally, on Bobcat. This is due to come out in sometime in 2009. I haven't seen a time frame for when in 2009 yet. Silverthorne, on the other hand is due out in 1H08. This gives Intel at least a 3Q lead and possibly a 7Q lead in this market. That should be plenty of time to establish themselves. The IP and "relationships" aren't worth much if the only offering belongs to your competitor.

Scientia from AMDZone said...

fujiyama

Torrenza - Available Today

All you need for this is either an add-on device that plugs into an AMD CPU socket or a card that plugs into an HTX slot along with an HTX motherboard. There are products of both types.

Trinity - Should be supported by VMware. This is either available now or soon.

Raiden - This is built on top of Trinity and also makes use of thin clients. I don't believe anything is ready with this yet.

AMD has made a number of tools and specifications available for these initiatives so there should be real products coming.

Scientia from AMDZone said...

intheknow

"Am I supposed to assume that up to this point they did not bother to utilize their assets"

You are missing the point. ATI's relationship with TSMC is an existing asset. AMD will use them in the future to make some Bobcat products. This is a change since even the Geodes were made in AMD's FABs.

"By comparison, when Intel restructured a while back"

You are talking about two different things. Asset Light is not about restructuring. It's about future products and some cap ex avoidance.

"That amounted to $2bil for the current year"

I don't think it would possible for AMD to trim like that. Intel is still not as lean as AMD.

"And I don't see anything in what you have posted that says whether or not they are going to outsource CPUs or not."

Are you still beating this dead horse?

Slide 149: GPU, Chipsets & CE Foundry

Slide 148: Shared Objective (TSMC) - High Performance Grahics/Chipset/SoC production on leading edge bulk technology.

Slide 147: Shared Objective (Chartered) - High yield foundry production of AMD microprocessors.

"Finally, on Bobcat. This is due to come out in sometime in 2009."

AMD already has products in place. Bobcat is the upgrade.

"Silverthorne, on the other hand is due out in 1H08."

Intel is starting from scratch. They are way behind AMD.

Slide 128: Over 200 Million Imageon processors shipped. Current customers include Motorola, Nokia, Panasonic, and Samsung.

Xilleon - Current customers include JVC, Mitsubishi, Samsung, Sharp, Sony, and Westinghouse.

What customers does Intel have for Silverthorne?

Ahmar Abbasi said...

AMD should be slideshow company of the year. We all know AMD has real skill with the power point now lets see some actual product from them........

Scientia from AMDZone said...

ringinheat

"AMD should be slideshow company of the year. We all know AMD has real skill with the power point now lets see some actual product from them........

I assume reviews will come at product launch which may be in August.

Scientia from AMDZone said...

fujiyama

"I'm talking about Turion X2 CPU for laptops where only 90nm are available."

It is difficult to tell with TL-56, TL-60, and TL-64 because these are available in both 90nm and 65nm versions.

However, HP currently offers a notebook with TL-58 which is only available as 65nm. This shows that 65nm X2 Turion is available.

"So find TL-66 CPU in Dell, Toshiba or HP offerings."

Dell mentions the TL-66 but it is not available yet. I haven't found anyplace that has the TL-66 yet.

Roborat, Ph.D said...

Scientia said: "lately I've seen people pushing the idea that AMD stuffed the channel in Q4 06 and this then led to the sharp drops in Q1 07. I can't say that I entirely blame these people for such speculation because I've been puzzled about this myself... The only explanation was that AMD had pulled off a near miracle and gotten 65nm ramped at FAB 36 in record time."

AMD specifically said that the saw "significantly lower demand, poor product mix and logistics problems in Q1'07. I don't understand why you insist on your own explanation and not even believe what AMD says. Just look at the inventory levels and how it shot up by $100M in Q1 which easily debunks your idea that the low volume was due to 65nm conversion. Nobody builds up inventory when their factory is supposedly converting.

Aguia said...

AMD should be slideshow company of the year.

13ringinheat,
and Intel can easily take the place of the on paper specs that does not work.

Nobody builds up inventory when their factory is supposedly converting.

What are you trying to say Roborat, Ph. D.?
They should not have products at all?
I think you always should have products on stock if you won’t be able to manufacture them for a while. It seams you forgot the huge inventory that Intel had one year ago.

Scientia from AMDZone said...

roborat

AMD has never admitted to stuffing the channel. Show me anything that supports this idea.

AMD's volume only increased slightly in Q4 compared to a large increase by Intel. This doesn't match with over-shippping. Secondly, the DDR based chips that had to be written off do support the idea that demand dropped in that particular area. This again would be consistent with a bad product mix.

Scientia from AMDZone said...

roborat

I just don't understand how the idea of stuffing the channel (over-supply) in Q4 fits with the reports of chip shortages in Q4.

For example, AMD talks about processor shortages.

AMD responds to processor shortage controversy.

How do you have shortages and oversupply at the same time? Are you claiming then that AMD only stuffed the channel in certain segments? Even if you are claiming this you should know that it isn't possible to oversupply to the degree that AMD's volume dropped in Q1.

It simply seems to be the case that AMD did not have enough of the stock that customers wanted in Q1. Q2 looks better in terms of volume but Q3 should see some financial improvement as well.

InTheKnow said...

I don't think it would possible for AMD to trim like that. Intel is still not as lean as AMD.

And here you have missed the point. Intel gave guidelines to measure the success of their efforts. I don't see any measurable metric from AMD for me to see that their plans are achieving the desired result. All I got was marketing doublespeak.

Are you still beating this dead horse?

Beating what dead horse? All I did was state that you hadn't provided evidence to support your claim. Now you have, thanks.

AMD already has products in place. Bobcat is the upgrade.

The implication being Intel does not? They currently have the McCaslin platform complete with products.

For an example see: http://www.digitimes.com/systems/a20070718PD215.html

Has Intel got a huge product line in this space? No. But what Intel is going after is a new product space and nobody has a lock on it. To try and claim that "Intel is starting from scratch. They are way behind AMD." is not an accurate evaluation of the situation.

For the record I understand this product space to be a market for full blow PC capability/internet connectivity in small form factors. These devices will be underpowered from the standpoint of compute capability compared to laptops and desktops, but will have all the functionality that most users want and need (i.e. email, IM, internet, etc.).

What customers does Intel have for Silverthorne?

From the above link:
In addition to Fujitsu, a number of companies, including Gigabyte Technology, BenQ, High Tech Computer (HTC), Averatec, Medion and China-based Hanwang, have all expressed their support for Intel's ultra mobile platform, according to market sources.

If they are interested in McCaslin, I'm certain they are will have an interest in Silverthorne, with its improved power and performance.

InTheKnow said...

Oh, you should add Samsung's Q1 Ultra to the list as well. I think you would have to count them as having potential to be a substantial volume customer.

Roborat, Ph.D said...

Scientia said: "AMD has never admitted to stuffing the channel. Show me anything that supports this idea."

Can you show me a company that admits to over supplying just to hide a massive decline in customer demand?

AMD's last two quarters, the massive dip followed by a recovery normally points to an even which occurred in Q4'07 when AMD tried to hide the shift in the market. If you think this theory is incorrect, that is fine. At least this theory is consistent with AMD's results and press statements. What is interesting now is your theory.

I think you owe your readers a better explanation why AMD's inventory shot up $100M when you allegedly said that their 65nm transition caused the massive decline in revenue, ASP and margins.

Scientia from AMDZone said...

intheknow

I think metrics at this point would be redundant. AMD's measure of success is how much less money they lose each quarter.

Thank you for the examples. However, they certainly don't seem to compare to what AMD has for Imageon and Xilleon. I think you would have to agree that Intel is behind.

The Bobcat chip would be useable for palm devices and maybe thin and light notebooks. Essentially, Bobcat combines AMD's current Geode and embedded MIPS processors.

Scientia from AMDZone said...

intheknow

Yes, Samsung could be big customer. So, Intel does have one.

Scientia from AMDZone said...

roborat

Let's follow your logic for a moment. You claim that:

1. AMD over-supplied the market in Q4.

2. AMD's volume dropped in Q1 because of too much inventory in the channel.

3. AMD's volume returned in Q2 because the inventory had been sold down by then.

At first glance this would make sense. However, it doesn't take into consideration that AMD's capacity has expanded. AMD gets about 40% more capacity from 65nm and FAB 36 is at about 75% capacity compared to 50% in 4Q. This means that FAB 36 has doubled in capacity which puts total production at 50% more than Q4.

Therefore: If AMD had actually been able to oversupply in Q4 then AMD's capacity today must greatly exceed its market demand. This makes no sense because if AMD had extra capacity in Q4 then it could have started scaling back FAB 30 in Q1 instead of 6 months later in Q3. This would have benfitted AMD both because of 200mm tooling sales and a greater ratio of 65nm chips which are cheaper to produce. So, let's try a scenario that fits both with production and observation.

Q4: AMD is short of chips. Thus while demand increases a lot in Q4 AMD's gross revenues only increase slightly. AMD is particulary short of 65nm since none are delivered until the end of November.

Q1: Having failed to deliver chips to some customers in Q4, AMD finds that these customers have bolted to Intel. AMD hopes to move more chips with other customers.

However, Intel has released L7400 and L7200 C2D mobile chips plus cheaper Allendale E4300 chips so AMD finds that demand for its extra chips has declined. AMD finds itself still short of more valuable 65nm chips and short of the new 6000+.

Q2: AMD still finds itself short of not only higher value 6000+ but the new BE chips as well. This is one reason why AMD's Gross Margin only rises slightly.

With additional capacity and more 65nm chips AMD is able to gain back some customers and its volume increases.

However the shift in demand from Q1 continues as Intel releases Allendale E4400 and the Celeron class Conroe L chips. With this additional produce pressure from Intel AMD is unable to move what are viewed as outdated DDR based chips. AMD's capacity is finally enough to meet customer demand. Announces it will begin reducing FAB 30.

Which of these scenarios actually matches?

Azmount Aryl said...

I can tell you a few things you might wanna know. in Q1 AMD were short on 6000+ chips. First, they were overly expansive and hard to get, then they got really cheap and still - hard to get.

In Q2 AMD were short on FX-70/72/74 (FX-70 got discontinued some time during Q2), I was really frustrated about that in particular. On one hand they got 6000+ models going in big volumes (3.0GHz) on the other - they are short on FX-74(3.0GHz) so that even i couldn't buy those.

As for my opinion on this shortage in Q1, well, i don't know for sure.
It could have been that they oversupplied channels in Q42006 but it also can be that they got dumped by Dell and didn't set prices right fast enough in Q12007 which led to mediocre unit sales for that quarter.

Christian H. said...

Something else is that R600 now has the top spots in 3DMark06 in XFire.

That was a good fing on that SPECfp test. I guess it has to be true and now people can stop saying Barcelona won't be competitive.

24% faster at a 330MHz clock speed deficit should shut up the naysayers and make it clear why they will be able to charge more for both Barcelona, Budapest and Kuma.

A little math will show that with the same amount of cores 10h (the actual processor family name) is about 40% faster at least in fp. This gibes with AMDs earlier numbers.

The test also means that anything below 3GHz will get kicked by a 2.5GHz Barc as we can assume a 500MHz lead (330MHz faster couldn't do it).

Kudos to AMDs engr team for delivering.

Christian H. said...

1. Torrenza, Raiden, Trinity - cannot buy them.
2. 65nm Turions X2 - cannot buy them
3. DTX from factor - again
4. HT 3.0 chipsets - ...



Raiden, Trinity are not products, they are initiatives. Torrenza was not due in 2007.

65nm Turions are being sold in HP laptopsand I believe Toshiba is also using TL-64 (anything above TL-60 is 65nm). Again, there is no fanfare when new systems become available. Most new Turions will ship with 690G which means extra time to qualify the chipset.

AMD has an HT3 desktop chipset RD790 which is more than likely what they ran the 3GHz Phenom demo with. None for server yet but they want the upgrade market to keep their current share.


It's no wonder I don't even bother commenting on these sites anymore. The horse is dead but it's still being beaten.

Sci, you should OVEREMPHASIZE THAT SPECfp link.

enumae said...

Christian M. Howell
I guess it has to be true and now people can stop saying Barcelona won't be competitive.

Considering that K8 was about 20% faster in 2P (4 cores) at the same clock, and AMD 4P vs Intel 2P (8 cores) AMD is about 45% at the same clock, or 25% faster with a 330MHz clock disadvantage.

I understand that AMD has been able to put 8 cores in a 2P setup while consuming less power (large performance/watt advantage in FP), but the absolute performance advantage in FP is about the same as it was with K8.

What we need to see is Integer performance.

24% faster at a 330MHz clock speed deficit should shut up the naysayers and make it clear why they will be able to charge more for both Barcelona, Budapest and Kuma.

Like I said above, we need Integer performance numbers.

If they are still behind in Integer Intel will continue to dictate prices.

abinstein said...

enumae -
"Considering that K8 was about 20% faster in 2P (4 cores) at the same clock,"

No, when running SPECfp 2006 on dual-socket dual-core, K8 is just about 10% faster than Core 2. In particular, look at this performance graph and compare the top green line (K8) and the violet line below it (Core 2).

It seem K10 extends the lead to almost 3 times that of K8.

"What we need to see is Integer performance."

This is funny. Most computationally intensive works are floating-point, with SETI and folding being the exceptions. Servers use a lot of integers but memory, I/O, and even context switch speeds are usually more important there. Other integer works on the desktop space do not require super faster computing, and K8 or even Turions are more than enough.

This is why AMD is showing SPECfp, because for most users who not only demand but also rely on high performance out there, floating point is what they need.

abinstein said...

Oh BTW crypto is another notable exception that needs integer performance.

Aguia said...

abinstein the true is that even current Opteron is not that far (or even ahead) of current new Xeon systems. Performance, power consuming, platform, processor design, ...

And all that with outdated 90nm manufacturing process and already more than 4 years old processor design.

It’s impossible for Barcelona to waste all this even if it is just one Quad core K8.

Desktop depends. Mobile even more. But with Intel increasing the TDP there it will not be a big issue.

enumae said...

Abinstein

Please look at SPEC...

AMD 2P = 52.5
Intel 2P = 43.1

AMD has an 20%+/- advantage.

This is funny.

I understand that you know more about this stuff than I do, but please stop the sarcasm.

PS: I am still waiting on a response from you about Intel needing 1600FSB on desktop for a 3.2GHz Penryn (last article), please post it here.

----------------------------------

Do you understand the comment that I was responding to was about AMD being able to price it's products higher because of an advantage in FP?

Considering that Intel was able to gain market share in servers and workstations while AMD held the FP advantage I believe that until AMD has an advantage in Integer Intel will control pricing.

Do you agree or disagree?

Unknown said...

Regarding the R600...
I looked at those slide and they didn't mention the R600 specifically.
Did they say it on the presentation?
I seem to remember Charlie or Fuad mentioning that the 65nm R600 was just part of an internal mole hunt. But then, AMD is about to release the new FireGL cards and the stream processors (Firestream... or what ever they call them), so it would make sense to get rid of the leakage problem of the 80nm version.

Axel said...

Since Scientia has a nasty habit of deleting / editing posts that don't jive with his worldview, I will be copying this post onto Roborat64's safe haven for blogger comments where I don't have to worry about censorship reminiscent of fascist times.

I could be wrong, but it's pretty clear to me that AMD not showing any benchmarks from the 3 GHz Phenom "demo" pretty much confirms that K10 will not compete with Penryn clock-for-clock. Cannibalization of current K8 sales is a non-argument. The vast majority of AMD's K8 sales are to OEMs, to customers who could care less about a Phenom demo and wouldn't know the difference between K10 and the original Pentium. Most of the enthusiasts who would care about the demo have already jumped onto the C2D / C2Q bandwagon anyway.

AMD would cannibalize an insignificant amount of sales if they showed off shockingly fast K10 benchmarks. OEM sales would be unaffected. Tech-savvy IT pros responsible for major server purchases aren't buying Opterons right now, they're hanging on for K10 benchmarks! It would be dumb for an IT pro to buy K8 when much better performance per watt is right around the corner. Enthusiasts have either already jumped to Intel or are waiting for K10 anyway. So exactly whose sales would be cannibalized?

No. The only logical explanation is that K10 isn't all that. It's not going to bring ASPs up enough to bring AMD back into the black. Inexpensive quad-core Penryns will be a far better value. Intel is apparently preparing a 45nm assault of massive proportions. According to TGDaily, in Q2 08 45nm will make up 50% of Intel's notebook CPU shipments!

http://www.tgdaily.com/content/view/33130/118/

Scientia from AMDZone said...

We only have two benchmarks. The SPECfp is a general purpose FP benchmark and K8 handles general purpose code better than C2D. K10 may have less of an advantage in full speed (tuned) SSE applications like encoding.

However, the VMware numbers should be about what you would get with Integer. If this is true then K10 will be faster than either Conroe or Penryn at the same clock. I have fairly good confidence that this is true because it also matches the earlier SPECint estimates. This would mean that AMD would be able to charge more for K10 on a strictly clock by clock basis.

Let's assume that AMD is able to deliver a 3.0Ghz quad core FX model in Q4. Presumably this would be priced higher than Intel's top quad core Kentsfield, the QX6850. However, AMD would probably price equally to Intel's top bin if Intel replaces it with a faster Penryn 3.2Ghz QX6950.

However, I don't know if AMD will be able to cover the full desktop range from 2.3 to 3.0Ghz in Q4. AMD may top out at 2.3, 2.4, or 2.5 in Q4 and not deliver the full range until Q1 08.

However, I'm certain that if AMD can deliver a 2.5Ghz desktop chip, its price will exceed that of the 2.66Ghz Kentsfield. It could even match that of the 3.0Ghz model.

2.3Ghz K10 will most likely match 2.66Ghz Kentsfield. I could also imagine 2.0Ghz being priced similar to 2.4Ghz Kentsfield.

Scientia from AMDZone said...

Björn

AMD officials twice stated that 65nm GPU chips are currently in production. I haven't seen anything that these could be other than R600.

The RD790 chipset in Q4 will be 55nm. And, the R700 series GPU replacement will not be out until 2008. It was originally thought that AMD would change R600 to 65nm and this was repeated in an interview with an ATI official. Fuad then decided later that this wasn't true and that AMD would skip to 55nm. Again, Fuad's second statement does not match what AMD officials said during the presentation.

Ho Ho said...

christian
"Something else is that R600 now has the top spots in 3DMark06 in XFire."

First, no it does not. Two 8800Ultras beat it just a few days after XT's did the record.

Secondly Single XT beats single GTX in 3D mark 2006. What does that mean? It means that R600 is good in 3D mark but not nearly as good in real world games.


"24% faster at a 330MHz clock speed deficit should shut up the naysayers and make it clear why they will be able to charge more for both Barcelona, Budapest and Kuma."

I'll try to make a nice table to see what exactly is going on with SpecFP and how do the results scale. SpecFP does show quite a speed difference against Xeon but comparing to older dualcore Opteron at same clock speed overall speed increase is not nearly as good as one would expect from twice the number of much better cores.

My non-checked guess is that even AMD direct connect architecture is bottlenecked in at least some of SpecFP benchmarks. I'll try to make a nice table of the data when I get time. I can't say when I do it as it takes a whole lot of time to compare hundreds of systems and thousands of benchmark results. I also have one hell of an interesting job so I spend most of my evenings doing that instead of prooving the claims I make here. Pe patient, I will show the results one day.


"AMD has an HT3 desktop chipset RD790 which is more than likely what they ran the 3GHz Phenom demo with"

I'm qurious, what benefit does HT3 have against HT1 on desktop?


"Sci, you should OVEREMPHASIZE THAT SPECfp link."

If he does then he should also compare it against older Opteron results.


enumae
"What we need to see is Integer performance."

That would indeed be more interesting as most programs depend more on integer throughput. Also SpecInt_rate doesn't depend on memory bandwidth as much as SpecFP thus it doesn't give as biased results.

Biased as in biased towards systems with better memory throughput leaving CPU performance on the background.


abinstein
"Other integer works on the desktop space do not require super faster computing, and K8 or even Turions are more than enough."

It doesn't matter that most people don't need much better HW than they already have. Most buyers just get what salesmen tell them to buy and they naturally try to sell as fast (==expensive) thing as possible.


"This is why AMD is showing SPECfp, because for most users who not only demand but also rely on high performance out there, floating point is what they need."

This is just probably just a fruit of my weird imagination but perhaps SpecFP doesn't show nearly as good results against faster clocked Xeon as SpecFP does. THat wouldn't look all that good to show to people, at least not yet.

Scientia from AMDZone said...

axel

"Since Scientia has a nasty habit of deleting / editing posts that don't jive with his worldview"

LOL. It's curious then that I haven't deleted roborat's comments; I must be slipping in my fascist rampage.

Axel, you can give any opinion you want as long as you don't flame. For example, if you start off a post with something like, "You're such a fanboy idiot . . ." then it will be deleted or edited.

"I could be wrong, but it's pretty clear to me that AMD not showing any benchmarks from the 3 GHz Phenom "demo" pretty much confirms that K10 will not compete with Penryn clock-for-clock."

How could you be wrong when you have Ed Stroligo at Overclockers agreeing with you? Ed's article is his typical low quality, shrill rant that looks like it was thrown together in ten minutes.

Ed has never gotten past his delusion that AMD and Intel should do marketing and product releases to cater to the 0.1% of the market that OC'ers represent. The truth is that the vast majority of computers buyers will never overclock or upgrade a cpu and don't care about getting a high score in SuperPi.

"Cannibalization of current K8 sales is a non-argument. . . . So exactly whose sales would be cannibalized?"

We know that Intel's sales were cannibalized in 2006. So now you would have to explain how it could happen to Intel but not happen to AMD.

"It's not going to bring ASPs up enough to bring AMD back into the black."

No, not until Q1 08. The ASP's for server chips should rise in Q3 and Q4 but AMD needs volume on the desktop and this won't be out until Q1 08.

"Inexpensive quad-core Penryns will be a far better value."

Then I assume you would agree that inexpensive K8 X2's seem to be a better value than Conroe for most buyers today.

I think Intel's quads probably will be a better value in Q4 but this is consistent with K8's being a better value now. I think it will take until Q1 08 for AMD's K10 chips to have enough volume to have a good price.

"Intel is apparently preparing a 45nm assault of massive proportions. According to TGDaily, in Q2 08 45nm will make up 50% of Intel's notebook CPU shipments!"

I guess I'm puzzled where the massive assault is. 50% 45nm for mobile in Q2 08 would be a reasonable ramp. And, it is good that Intel will have 45nm chips for notebooks because it will be up against AMD's new mobile processor and chipset. And, it looks like Intel will need 45nm chips to compete with K10 in desktop and server.

Scientia from AMDZone said...

ho ho

That isn't necessary. ATI officials specifically stated that XT was only intended to compete with GT and not GTX. They stated that ATI counts on Crossfire to compete with GTX and higher. In other words, ATI did not release an XTX model that would have matched GTX.

SPECfp. When I looked at SPECfp it seemed to me that K10 was getting about 15% higher performance than K8. However, SPECfp is not a full speed SSE application. It is more general purpose computation which is why K8 can have a higher speed than C2D. For a full speed comparison you'll need something that has loops of almost pure SSE computation.

I agree that SPECint would be a better predictor of general speed. AMD was claiming 20% higher Integer than C2D at the same clock. Since the SPECfp numbers seem to match (51% higher actual versus 50% simulated) and because the VMware numbers are similar (34% greater IPC than K8) we can probably accept the 20% figure.

Axel said...

Scientia

Axel, you can give any opinion you want as long as you don't flame.

Several of my comments over the past couple months have either been deleted wholesale or edited here and I have never flamed. My latest comment about fascism is the closest I've come to flaming.

And I have seen countless comments by others mysteriously disappear or re-appear in an altered form. You can go back and check your repository, then you'll see the truth about what you've been deleting. A good half of the deleted / edited comments have been germaine to the topic at hand and offered meaningful insight.

Therefore I will continue to copy my posts to Roborat's blog (including this one) until I see a positive change towards respect for freedom of speech in your moderation. In fact I post it there first before hitting the publish button here. I've already been burned several times and it's a pain to lose a comment that took a good ten or fifteen minutes to compose. Trust me, you are deleting more than flamer comments. Just take a look at your records.

We know that Intel's sales were cannibalized in 2006. So now you would have to explain how it could happen to Intel but not happen to AMD.

Can you supply a link proving that Intel's P4 sales were cannibalized specifically due to the C2D benchmarks that were shown during the Spring IDF? I don't think you'll be able to convincingly show any such thing. Intel was losing ground in P4 sales and marketshare well before IDF as K8 was rapidly gaining mindshare and acceptance. I would venture that only IT pros and enthusiasts were really interested in the IDF benchmarks. Around that time, both of these groups were far more likely to purchase AMD K8s than the inefficient Intel P4s.

No. What the IDF benchmarks and circulating engineering samples accomplished was more to cannibalize *K8 sales* than *P4 sales*. Tens of thousands of Intel users who were ready to go AMD after running P4 furnaces for a couple years or AMD users who were ready to upgrade to the next Opteron or Athlon64 X2 saw the IDF benchmarks, stopped, and decided to wait. Then they checked out xtremesystems and saw how fast these ES Conroes were. None of these people was about to buy a P4. Why would people in the know buy a P4 when a next generation product is around the corner? So the benchmarks simply changed their original decision from
- Buy K8
to
- Wait FOR C2D.
Cannibalization of P4 sales due to pre-release benchmarks doesn't even come into the picture.

OEM sales in Q2 06? Unaffected by the IDF benchmarks and ES Conroes floating around. The average Joe didn't see the benchmarks, didn't care, doesn't know what IDF even is. He bought the K8 or P4 that he was going to buy anyway, decision completely unaffected by the benchmark hoopla.
Now I'll grant you that P4 sales were cannibalized AFTER the release of Conroe. But that's not the issue, we're specifically discussing AMD's reluctance to share pre-release performance information.

By showing impressive benchmarks now, AMD would at least forestall purchases of C2D/C2Q products even if their own K8 sales would not improve. But of course AMD can't show impressive benchmarks because, well, they're not that impressive.

Scientia from AMDZone said...

SPECfp. There is a really good example of a memory bottleneck with a Dell Xeon 5160 and a Fujitsu Xeon 3060. Both are single socket/dual core processors and have 8BG's of memory.

Xeon 5160 3.0Ghz: 24.3
Xeon 3060 2.4Ghz: 24.3
Opt. 2216 2.4Ghz: 22.5

5160 is clocked higher and has a faster 1333Mhz FSB with faster 667Mhz FBDIMM while 3060 has a 1066Mhz FSB and slower 525Mhz memory. However, 5160 only uses a single FBDIMM while 3060 uses two DDR2 DIMMs. This gives 3060 an advantage in bandwidth.

We also have a Fujitsu Opteron 2.4Ghz 2216 which is also single socket/dual core and is not limited by memory bandwidth. It's SPECfp scores is slightly less than 2.4Ghz C2D.

These are ranked in order of speed for each test. NBI = Not Bandwidth Intensive.

bwaves: 2216, 3060, 5160
games: 5160, 3060, 2216 NBI
milc: 2216, 5160, 3060
zeusmp: 2116, 5160, 3060
gromacs: 5160, 3060, 2216 NBI
cactus: 5160, 3060, 2216 NBI
leslie: 2216, 3060, 5160
namd: 5160, 3060, 2116 NBI
deal: 3050, 2216, 5160
soplex: 3050, 2216, 5160
povray: 5160, 3060, 2216 NBI
calculix: 5160, 3060, 2216 NBI
gems: 2216, 5160, 3060
tonto: 2216, 3060, 5160
lbm: 2216, 5160, 3060
wrf: 5160, 2216, 3060 almost 3 way tie
sphinx: 3060, 2216, 5160

It is clear from these tests that about half are bandwidth intensive. We could better judge K10 if we had a breakdown of the SPECfp subtests.

Ho Ho said...

scientia
"ATI officials specifically stated that XT was only intended to compete with GT and not GTX"

I know that, I was just saying that 3D mark doesn't mean almost anything when it comes to real games. Kind of like Spec, it depends on what you are interested in.


"When I looked at SPECfp it seemed to me that K10 was getting about 15% higher performance than K8"

I wish it was that simple.
Compare those two systems (sorry, no time for bigger analysis):
2GHz
3GHz.

Same architectre, nearly the same configuration (nothing should have too big effect on results), different clocks.

Going from 3GHz down to 2GHz, an dencrease in clock speed of 33%, we get roughly 22.5% worse result. If clock speed would be the only thing affecting the benchmark then we should see roughly 22.5% speed difference in every sub-benchmark. Same should go if memory bandwidth is not an issue.

Now lets put all the sub-benchmarks side-by-side:
chart1

Well, well, well. What do we have here. Not all things scale equally. Let's see how well they do scale:
chart2

As you can clearly see things are quite odd in some places. Scaling ranges from ideal 33% all the way down to negaive -5% scaling. As was said before average comes down to around 22.5%.

Now question is why exactly do some benchmarks scale worse? It can't be CPU execution efficiency as it is same for both of them. To me it seems as only thing that can have so big effect on the results is memory system as this one didn't scale with CPU clock speed. Question is if it is raw bandwidth, latency or something else.


Now let's try to compare the results with K10. K10 with 8 2GHz cores has a score of 69.5. With half the number of K8 cores AMD scored 40.7. Increase core count by 100%, gain 71%. Kind of small gain for more effective core, especially when considering that K8 in 2P at 2GHz is very little ahead of same specced dualcore Core2. Comparing to 3GHz K8 (same power envelope according to VMWare slide) you'll gain only 32%.

In fact I wouldn't be surprised to see the difference between K10 and Core2 to grow smaller with increasing both of their clock speed, mostly thanks to Core2 being able to use its larger cacche (twice the capacity per core vs K10).

I don't currently have time to do similar charts about Core2 or other CPUs but I bet we'll see similar results. With Core2 (and non-availiable K10 result) I excpect to see much higher variation with those bencmarks that scale linearly with clock speed to keep on scaling that well and others to lag behind even more than with K8.

To get an accurate picture of the things much more systems should be analyzed. Currently I only took two of the highest ranking ones in their respective categories. As I said I have no time to create an automated system that would do the work for me. I might do it in a few days or couple of weeks but don't count on me too hard. It isn't too hard to make something yourself and you'd probably get results sooner if you did it yourself.

If I made some mistake somewhere I'd be glad to hear about it. I wouldn't like to write a bunch of code just to find later that it is buggy.



because the VMware numbers are similar (34% greater IPC than K8) we can probably accept the 20% figure."

You'll have to take into account that VMWare also benefits from virtualization improvements in K10. If what AMD said is true then those improvements are substantial.

Another interesting thing to note on that slide is that they say Opteron gets 79% performance increase at same power. They compare against 3GHz dualcore opteron. Unfortunately they don't tell which one exactly but if it is DDR2 one then keep in mind it has TDP of at least 119.2W, unless they compare it with 1P one at 125W.


"These are ranked in order of speed for each test. NBI = Not Bandwidth Intensive."

Interesting thing is when I look at the scaling chart then it doesn't match with what I found. SOmeone should make similar chart for Core2.


As a movie title once said, "Analyze That" :)

Scientia from AMDZone said...

I don't want to spend a lot of time discussing the way I manage the comments on this blog. Comments that have been deleted or edited lately have contained insulting language.

There are posters though who are used to using insulting language when posting on other forums or blogs. You only have to look at the comments in Sharikou's blog to see examples of this. Since I can't have 100 different posters all deciding differently what is or is not insulting, the standard has to be mine. I'm sure there are many who don't think that calling someone a fanboy or a liar should count as an insult but that is the standard.

"Can you supply a link proving that Intel's P4 sales were cannibalized specifically due to the C2D benchmarks"

No, but it's a reasonable assumption based on both Intel's volume and price changes.

"What the IDF benchmarks and circulating engineering samples accomplished was more to cannibalize *K8 sales* than *P4 sales* . . . Tens of thousands of Intel users who were ready to go AMD"

Axel, I know you are trying to make a good argument. Okay. But this is more speculative because it is based on what people might have been thinking. It also makes no sense. Why would 10,000 users make any difference to a company that was making 1 Million chips a week?

Let's look at the volume share from Q4 05 to Q4 06:

Q4 05: AMD 22.1%, Intel 77.9%
Q1 06: AMD 22.1%, Intel 77.9%
Q2 06: AMD 22.9%, Intel 77.1%
Q3 06: AMD 23.4%, Intel 76.6%
Q4 06: AMD 25.4%, Intel 74.6%

This doesn't seem to match the idea that Intel reduced AMD's sales with its C2D announcements. In fact, since AMD was short of chips in Q4 06 their volume share could have been even higher.

"But of course AMD can't show impressive benchmarks because, well, they're not that impressive."

Do you know what that comment reminds me of? Way back in 2003 I said that Prescott would have a 30 stage pipeline and people told me I was crazy because Intel wouldn't add 10 stages to P4's already long 20 stage pipeline. When Prescott was actually released though it had 31 stages and Intel had added 11.

Recently, I said that AMD had K10's working at 2.4Ghz and again was told that I was crazy because Anand and Fuad claimed that AMD didn't have anything faster than 1.8 or 2.0Ghz at the most. But just a short time later AMD demoed 3.0Ghz. So who do you think was right?

At 2.0Ghz K10 is not going to beat a 3.0Ghz C2D. So, in that sense, K10's scores will not be that impressive until the speeds are higher. However, the per clock speed should be good.

Finally, the part you left out in your argument is that Barcelona will be officially released in either August or September. This gives less than two months (but probably more like 1 month) for any testing information released now about K10 to have any effect before the NDA is lifted and reviews appear. After we factor in how much this would reduce coverage of the official launch I don't think there would be any gain.

Axel said...

Scientia

Let's look at the volume share from Q4 05 to Q4 06:

You missed my point. Overall volume share numbers are meaningless because OEM sales dominate those figures. My point is that it's specifically the sales to the IT professionals & enthusiasts that were affected by the Conroe pre-release performance disclosures. You will see that server share turned around to Intel far quicker than the overall marketshare figures:
http://news.com.com/Intel+ups+server+share,+but+AMD+wins+in+PCs/2100-1006_3-6154735.html
That's because it's the IT pros that drive those sales. And the same probably goes for enthusiast share, though I'm not sure there's a way to break that out separately from public domain data.

Like I said, you won't find really data to support your case because the "IDF effect" is completely swamped by the OEM sales and difficult to really discern. But it's the only logical conclusion. I've already shown by logic how P4 revenue cannibalization due to IDF doesn't make sense.

I'll take a chance that you won't delete / edit this post and not copy it to Roborat's blog. I've noticed that you've been less heavy-handed recently. That's good progress.

Scientia from AMDZone said...

Oh, one correction on my SPECfp comparison. It should say that the 3060 has a memory advantage because 5160 is using 4 ranks of FBDIMM which adds latency with each rank while 3060 is using regular DDR2.

ho ho

The two systems are very similar except that the slower system has twice as much memory. Also, the tests are run at different User priority levels. It's interesting though because all of the subtests that are not bandwidth intensive show perfect or nearly perfect scaling.

lex said...

You fanbois don't get it do ya. You can debate benchmarks, FSB, IPC, power efficiencies blah blah blah but the bottom line is AMD can't win this long-term battle.

There most recent dog and pony show with hundred + pretty green PowerPoint doesn't change it one bit and here is why.

AMD today must compete at the platform level. As such they must compete like INTEL providing both the CPU but also the chipset and rest of the platform. The additional burden for AMD is that the x 86 businesses looks the same top to bottom. The only thing separating the top end from the bottom end is performance and price. If you don't or can't compete in a certain market you get kicked to the next lower market with commensurate exponentially higher volumes and lower margins. SO in the end you can't compete in just one segment you are competing in all segments.

Thus AMD was forced to go into huge debt to absorb ATI. Now they are in process of digesting, assimilating and hopefully getting their first generation of integrated platforms out to compete with INTEL across all the product lines. The naive, blind AMD faithful say just wait when Barcelona finally comes, when Phenom comes, when Fusion comes you'll see AMD rise again. That is simply not the whole story. I have full confidence Barcelona will finally come and someday get to 2.6 GHZ and run circles around Penrym. Phenom will arrive sometime in 2008 in volumes and also run give Penrym a run for the money. This doesn't end with this but you have to look out and step back to see the bigger picture.

Here is why it won't work. If the goal is to compete, AMD will compete. If the goal is to make money and have a sustainable roadmap and be competitive quarter in quarter out and be profitable so you can continue to compete AMD doesn't have a chance.

To have a sustainable business you need silicon technology, design/product technology, manufacturing scale and expertise.

Very simply if you lack any of these you simply can't sustain the level production volumes, levels of investment to generate required future technology, manufacturing and design/product cadence to sustain profitability for the full technology cycle across multiple segment lines to sustain the investment you need. Its a huge cycle, once you have it it continues to make you stronger and once you fall out its almost impossible to get it back.


So sure Barcelona will be competitive, so will Phenom. The larger problem is that with one fab running 1 year behind technology that is slower then INTEL's comparable technology will hinder whatever small advantage AMD might gain in their design cadence. That combined with their smaller manufacturing scale guarantees them to be forever handicapped.

Thus AMD has competitive in only one of the three areas they must be competitive most of the time. Somehow people think that INTEL's disastrous Netburst decision that resulted in the appearance of AMD superiority for a good part of two years was somehow some a monopoly AMD had on design or that design can win over superior manufacturing and silicon is simply not true. If Penrym, Nehalem, Westmere, and Sandybridge are correctly positioned AMD simply won't have any competitive advantage in design and not having competitive silicon nor the required leading edge capacity guarantees them to be losing money 3/4ths of the time.

How you sustain a business losing hundreds of millions of dollars for 3/4 quarters and possibly making a few hundred in the remaining one without addressing this is just silly. If this is so obvious you have to ask why does AMD continue? That is easy too! You have to remember that Hector, Henri, Dirk and Jerry simply can't step back to see this. They have spent their whole life fighting INTEL. They simply can't say it can't be done and we need to find a way not to compete. It will take an outsider ( like Lou Gerstner at IBM, Hurd at HP ) to stop and see the real landscape.

Look today at AMD, great plans, but inferior silicon, miniscule manufacturing and equal design. They can win a round or two but simply can't win the war. Nor do they have what it takes to invest to win the war. Then need tens of billions of dollars to invest in silicon and manufacturing to get to 30+% to get the economies required in this business.


Its like Boise state, got a few jocks, a great game plan, and a few tricks. They can win the Fiesta bowl and make Oklahoma look like chumps for one game. But put them in the BIg12 and have to play Texas, A&M, Tech, Nebraska, Oklahoma. They will be nothing but a B12 cellar dweller. AMD is similar, they haven't upgraded their technology, nor their manufacturing and thus they are going to be history. It takes years if not a decade to create the infrastructure and critical mass to be a powerhouse team in collegian football, the same in x86 and AMD doesn't have it.

Since the last challenge was about some details I can line up some numbers and they don't look pretty for AMD.
AMD's gross margins will bounce probably between a low in the 30s to a high in the mid 40's. So figure that gives them what at best between 2-4 billion in free cash. With that they need to catch up on 65nm, invest in a 45nm pullin, and start 32nm development. Oh out of the 2-4 billion also invest in manufacturing capacity for 45nm. Oh also fund all the design cost for those initatives/platforms on them 150+ PowerPoint foils.

Now compare that to INTEL with GMs between 40-50 percent on 30+ billion leaves them with 12-15 billion in cash to spend on the exact same initatives. Assume INTEL has to outspend them 4:1 for fab and INTEL still makes 5-6 billion profit, invests in same cadence of manufacturing, R&D while AMD sucks billion dollar losses trying to stay even, let along catch up on technology or grow manufacturing capacity. Sorry math doesn't add up. They are like Boise state playing with smoke and mirrors. They can win one game, but for the full season they are like I already said a cellar dweller and their stock will be below 10 bucks by end of 2008 when Nehalem and 45nm crush them!

What did I miss? Maybe INTEL's silicon R&d is bust, maybe Penrym, Nehalem, Westmere, Sandbridge is a bust? Maybe a scud missile destroys the fab in Israel and an electrical brownout takes out Arizona? Sorry, if AMD was to plan for INTEL success and expect to still be profitable the only logical and reasonable strategy is to find a niche where you don't compete with INTEL to make a little profit. You simply can't go head to head.

Scientia from AMDZone said...

axel

The server volume share shows that AMD started sliding in servers in Q4. That makes sense because Woodcrest was released first.

As I pointed out earlier, 10,000 enthusiasts won't make any difference in revenue. That would only be about 1/100 of 1% of AMD's annual volume. That number of sales would not be detectable even if they all bought AMD processors in the same week.

Axel said...

Scientia

That number of sales would not be detectable even if they all bought AMD processors in the same week.

We've gone off tangent. Let's return to the original question that started this whole thing:

Can you show how Intel's Conroe performance preview at the Spring 2006 IDF cannibalized P4 sales *during the four months before release*?

Remember it is this four-month interim period that we are focusing on, as the entire reason this question came up is people trying to figure out AMD's true motive behind keeping K10 performance under wraps for so long now. Stroligo added more fuel to the fire but his conclusion is actually more logical than the silly rumor out there that AMD are withholding such information for "competitive reasons" even as they lost $2 billion in a single year.

So you still haven't provided a good sound argument for your theory that pre-release Conroe benchmarks cannibalized P4 sales during that four month period. Whereas I have provided a fairly simple and logical argument for why this doesn't make sense:

1. OEM sales are not affected by benchmark hype.

2. The great majority of people who tuned in to the IDF benchmarks were of the tech savvy type who also recognized that K8 was a superior processor in almost every way to the P4.

3. Hence it doesn't make sense that these people would be contemplating the purchase of a P4 at that time, as they know better. They would either buy K8 or wait for Conroe.

4. Therefore P4 sales were not cannibalized by the pre-release benchmarks, as those P4 sales wouldn't have happened anyway.

5. Hence this argument doesn't make sense in the context of Barcelona either.

6. Therefore AMD must be covering up Barcelona's true performance because it's truly too little too late.

7. QED

Scientia from AMDZone said...

lex

"AMD today must compete at the platform level."

True which is why they bought ATI.

"SO in the end you can't compete in just one segment you are competing in all segments."

It should be obvious that AMD has been competing in all segments since it released Opteron and Turion.

"Thus AMD was forced to go into huge debt to absorb ATI."

Quite true.

"I have full confidence Barcelona will finally come and someday get to 2.6 GHZ and run circles around Penrym."

Well, I don't know about running circles but it should be competitive.

"Phenom will arrive sometime in 2008 in volumes"

Q1 08.

"The larger problem is that with one fab running 1 year behind technology that is slower then INTEL's comparable technology"

I'm not sure what you are trying to say. FAB 36 will probably be 7-9 months behind in 45nm. We don't know yet if this will be slower.

"That combined with their smaller manufacturing scale"

Not exactly. There are both advantages and disadvantages to having more volume. AMD should have a much easier time ramping FAB 38 than it did ramping FAB 36.

"If Penrym, Nehalem, Westmere, and Sandybridge are correctly positioned AMD simply won't have any competitive advantage"

I would say that Intel will lose all of its mobile advantage in 2008 against AMD's new mobile platform.

Desktop should be neck and neck.

Server is interesting though. Intel could gain a temporary advantage in 2H 08 if Nehalem is good before Bulldozer is released in 2009.

I can't really speculate out any further than that. I don't really have enough information about 32nm or plans into 2010.

"How you sustain a business losing hundreds of millions of dollars for 3/4 quarters"

AMD has spent a lot higher percentage of its income for capital expenditure. Adding a second FAB was a big step and purchasing ATI was another big step. Your mistake is in seeing these as continuous expenses rather than temporary expenses.

And, Intel wasn't always like this. Intel only started making chipsets in 1992 and introduced the PCI standard in 1993. However, it came together in 1995 when Intel introduced the ATX standard and released Pentium Pro. Intel replaced IBM as the standards maker for the PC and was finally taken seriously as a server chip maker. This is also when Intel began producing motherboards which allowed it to make more money at both IBM's and Compaq's expense.

AMD only released its first non-Intel K5 design the same year, 1995, but didn't produce a server chip for another six years. AMD is no longer six years behind.

"miniscule manufacturing"

Do you have trouble with scale? AMD produced about 65 million chips in 2006 and plans to be at 100 million in 2008.

"Then need tens of billions of dollars to invest in silicon and manufacturing to get to 30+% to get the economies required in this business."

What if I told you that 100 million chips in 2008 will be 30%. It remains to be seen if AMD will build a 3rd FAB. However, this would only cost around $3 Billion for the FAB itself and another $2 Billion for tooling. And, you need to understand that this cost would be spread over four years.

"They will be nothing but a B12 cellar dweller."

AMD's only competitor in processor manufacturing is Intel.

"AMD is similar, they haven't upgraded their technology, nor their manufacturing"

AMD constantly upgrades both. They are running 45nm chips today. AMD is still ramping FAB 36 which won't reach maximum until mid 2008. AMD is also now removing the old 200mm tooling in FAB 30. It will be replaced with 300mm tooling.

"AMD's gross margins will bounce probably between a low in the 30s to a high in the mid 40's."

No, you don't understand. AMD's gross margins only dropped to 28% in Q1 because it was still producing a lot of chips on the less cost effective 200mm tooling in FAB 30. This situation will never happen again.

Secondly, AMD was delayed in moving to 65nm because it didn't want to install new but obsolete 65nm 200mm tooling in FAB 30 and had to wait until FAB 36 was built. This scenario will never repeat.

ATI lost a lot of Intel orders when it was bought by AMD. This will not repeat. However, ATI's chipset is gaining quite a bit of popularity so it should be able to turn a profit soon. A better estimate would be 40-55%.

"they need to catch up on 65nm"

No. The 65nm R&D phase is over. It's now just a matter of tweaking production.

"invest in a 45nm pullin"

AMD already has a 45nm immersion scanner in operation. Test wafers are running now and we should see a Shanghai tapeout soon.

"and start 32nm development."

Start? AMD has been doing research on 32nm for some time.

"also invest in manufacturing capacity for 45nm."

That's already in the 2007 budget as $1.8 Billion.

"Now compare that to INTEL with GMs between 40-50 percent on 30+ billion leaves them with 12-15"

Intel spends about 3X as much as AMD on R&D but it has to fund Itanium development and all of the process development itself while AMD shares process development costs. Unless Intel increases R&D spending it has no advantage.

Intel does not have a chipset comparable to 690G and the replacement RD790 chipset will be out in Q4. Intel will need to increase development spending on chipsets and motherboards if it wants to catch up. Further, as Intel tries to develop discrete graphic solutions its costs will increase even more. I'm also not sure how Intel is going to respond to AMD's new mobile platform. What this means is that Intel's previous lead with Centrino is rapidly vanishing and there is no indication that it can maintain its position in 2008. This could easily mean large losses of mobile share.

Scientia from AMDZone said...

axel

"1. OEM sales are not affected by benchmark hype."

I'm not quite sure what you are trying to say. OEM's will get information long before consumers do. They got sample K10's some time ago. They will also get the newest chips for production before these are actually announced.

"2. The great majority of people who tuned in to the IDF benchmarks"

This was only a small fraction of consumers.

"3. Hence it doesn't make sense that these people would be contemplating the purchase of a P4"

Again, these people really don't matter.

"5. Hence this argument doesn't make sense in the context of Barcelona either."

Perhaps, but I assume you've noticed a few Core 2 Duo commercials on tv. AMD doesn't have the budget for large ad campaigns like this. So, it makes the most sense to release all the information at once for maximum effect.

Ho Ho said...

scientia
"But just a short time later AMD demoed 3.0Ghz."

AMD "demo" has been running two programs at extremely low clocks and both benchmarks gave some questionable results. That 3GHz box was sitting in inaccessible place (doing nothing but running OS?) and nobody outside AMD could touch it. Why is that? AMD should learn from Intel and let journalists play with their machines, it would increase their credibility a lot.

Also as you probably saw from the slides AMD doesn't expect to reach 2.6GHz on quadcores before Q2 next year. Almost a half a year after release. Will it be at 3GHz by next Christmas when Nehalem comes out?



"The two systems are very similar except that the slower system has twice as much memory."

How much memory does SpecFP_rate need to run 4 copies of itself? My guess is a lot less than either of those systems had meaning memory capacity has neglible effect. Perhaps only 2GHz one had slightly less performance (sub 2% difference) thanks to having more RAM sticks for CPUs. I guess we'll see better after I or someone else makes a full analysis of Spec data.


"Also, the tests are run at different User priority levels."

No they are not. Runlevel != priority level.

Their difference was that one had no networking (3GHz), other had networking (2GHz), neither had GUI. How much will networking support decrease benchmark performance?

All the differences you pointed out should give slight but most likely neglible speed decrease for the slower system. What do you think how much could have the results changed if both systems were configured exactly the same?


"It's interesting though because all of the subtests that are not bandwidth intensive show perfect or nearly perfect scaling."

Would you like to take a guess why there are some benchmarks scaling in between perfect an nearly no speed increase? What kept them scaling perfectly with clock speed increase?

Also, abinstein, do you still think that K10 won't be bottlenecked in SpecFP as you so politely said to me in sharikous blog?

Though I wonder why didn't you make any more comments on what I said. Do you agree or disagree with what I said about scaling and did I got the numbers correct?

I'd also like to hear comments about the ~32% speed gain at same power envelope. I doubt anyone likes to increase their CPUs to much more expensive ones to gain so little. Sure, there are people who will upgrade their whole systems and they won't care if those first CPUs are not all that great compared to the old ones. Though if some has lower clocked and less power hungry K8 servers they will have to either get the more expensive energy efficient models* or upgrade their cooling and possibly power supply systems.

*) Has anyone got any idea what wattage will the initial CPUs be? Perhaps all three ranges at 120, 95 and 68W with last two at lower clock speed? That VMWare slide makes me think that normal 2GHz ones will be at nearly 120W if not 125W as AMD doesn't have lower wattage 3GHz DDR2 Opterons. For DDR1 it has singlecores at 92.6W and it can't be comparing against those. Perhaps when AMD said you can use same cooling and power to cool down those Barcelonas they meant you can replace the 120W SE models that have been at nearly 120W at 2.8 and 3GHz. As I said VMWare slide backs up this theory.

Another thing I'd like to know is where did you got that 15% IPC increase compared to K8 in SpecFP. If only using Spec results then there is no way to know how much better IPC does it have as it scales awfully not even getting twice the speed of same clocked dualcore. Is it a good guess you took that number out of thin air without even comparing it to other results?


"As I pointed out earlier, 10,000 enthusiasts won't make any difference in revenue."

It might if significant portion of those make decisions for HW purchases in big companies.

Ho Ho said...

scientia
"FAB 36 will probably be 7-9 months behind in 45nm. We don't know yet if this will be slower."

You do remember that Intel doesn't ramp before it has squeezed out almost everything from the node and AMD releases a bit earlier and does continous improvements on the transistors? This is why it took Intel that long to get their power usage down, it was quite low already at the time they introduced the chips. I wouldn't say that at start of ramping their quality is comparable. It takes months for AMD to reach the level Intel releases its chips at.


"I would say that Intel will lose all of its mobile advantage in 2008 against AMD's new mobile platform."

You probably meant to say all expect performance and brand name ;)

Also Intel expects to have half of its mobile CPUs at 45nm in Q2 next year. Quite a fast ramp, I should say. Especially considering how big market share mobile PCs have.


"Desktop should be neck and neck."

Depends highly on raw clock speeds, Phenom and Penryn performance. We know almost nothing about any of those three yet. We can only assume that Phenom is faster than K8 and Penryn is faster than Core2 (not just clock speed).


"Your mistake is in seeing these as continuous expenses rather than temporary expenses."

It is continous expenditure to pay back those depts. How many years will it take and how much per quarter does it have to pay to get rid of those loans?


"Do you have trouble with scale? AMD produced about 65 million chips in 2006 and plans to be at 100 million in 2008."

Do you have any numbers for Intel or the whole market?


"What if I told you that 100 million chips in 2008 will be 30%."

I would say it is quite difficult to sell them.


"AMD's only competitor in processor manufacturing is Intel."

Power is also quite nice but probably only competes with Itanium as AMD doesn't have something in that big scale and it won't have for at least 1.5 years.


"They are running 45nm chips today"

Didn't Intel show their 45nm CPUs just a little after last Christmas running at rather high frequencies running some benchmarks?


"AMD's gross margins only dropped to 28% in Q1 because it was still producing a lot of chips on the less cost effective 200mm tooling in FAB 30."

How high were the margins in Q2 and how high will they rise in Q3? Above mid-40's? Somehow I doubt they recover that fast.


"This situation [low cross margins] will never happen again."

I take you don't think that continous price cuts throughout Q2 and beginning of Q3 doesn't affect AMD margins a lot.


"ATI lost a lot of Intel orders when it was bought by AMD. This will not repeat"

Intel orders are huge part of the market share it will never get back though.


"Intel spends about 3X as much as AMD on R&D but it has to fund Itanium development and all of the process development itself while AMD shares process development costs. Unless Intel increases R&D spending it has no advantage."

Intel also has some nifty things as silicon optics and massively parallel research chips and software that AMD probably doesn't have. What do you think, can copper wire based HT with lots of wires going from CPU to motherboard can compete with silicon optics?


"Intel does not have a chipset comparable to 690G and the replacement RD790 chipset will be out in Q4. Intel will need to increase development spending on chipsets and motherboards if it wants to catch up"

It is probably just my fault of not doing enough research but what exactly is so special about those chipsets?


"Intel will need to increase development spending on chipsets and motherboards if it wants to catch up."

This will probably ansvered by previous question but what exactly does Intel need to catch up with?


"Further, as Intel tries to develop discrete graphic solutions its costs will increase even more"

Does it show anywhere? Also AMD now also develops discrete graphics together with all sorts of other things like Fusion, GPGPU chips and other things. Intel discrete graphics seems to be general purpouse chip that suits well for many things including GPGPU and fusion-like things in addition to just rasterizing and ray tracing. Basically what I'm trying to say is that Intel does much more than just develops CPUs, chipsets and graphics. Adding a new branch doesn't affect it nearly as much as AMD who doesn't develop as many things.


". What this means is that Intel's previous lead with Centrino is rapidly vanishing and there is no indication that it can maintain its position in 2008"

What exactly tells us that AMD can gain any marketshare with its new platform? Sure, it does have better power managment than before and possibly slightly higher performance than regular K8's but Intel has 45nm dual and quadcore Penryns to put against it with considerably higher performance. Also it already does have quite nice power usage, 45nm will only make it better.



As for advertising, we don't have Intel or AMD advertising their products in Estonia. We do have retailers advertising their products, though and they seem to show Intel products much more than AMD. As usual main message is more performance for your money. There were quite a lot of AMD commercials just a year ago but not nearly as much any more.

Ho Ho said...

As I already had the template from last night I threw in two more systems:

2K 2GHz Core2
2P 3GHz Core2

This times I managed to find two equal systems and both of them had the highest result in their categories. Here are the updated charts:

Scores
Scaling

As I said before Core2 scales linearly in non-bottlenecked benchmarks and worse than K8 in the bottlenecked ones thanks to less system bandwidth.

Of the 17 benchmak K8 has 8 in which it doesnt't have bandwidth problems. Core2 on the other hand only has 5 of such benchmarks.

In 3 benchmarks AMD had considerable scaling problems and a bit less problems in three more. Core2 had big problems with 7 benchmarks and less problems with 3.


What is left to do is to compare Core2 quads and perfect scaling (SpecFP non-rate*number of cores) and see how will the benchmarks act then.


My current conclusion is that SpecFP_rate is good at measuring memory intensive performance but not as good with measuring CPU calculation speed.

E.g fastest SpecFP score is achieved by 4.7GHz Power6 running at 4.7GHz and having 4MB L2 per and 32MB of L3 cache per chip achieving 22.4 points. 3GHz Core2 gets 17.7 meaning it isn't too much behind it. Now going to 2P/4core we have Power leading with 115 and Core2 way behind with only 45.5 points. Massive memory bandwidth and cache is pretty much the only thing that counts in achieving high SpecFP_rate scores.

One thing I noticed is that neither Core2 nor K8 can run fluid dynamics that well. Also it is kind of weird that from the 17 beenchmarks 4 were calculating fluid dynamics and only one of those didn't show massive scaling problems on both platforms.


I wish I had more time to put together some nice table with all the interesting things and not just x86 CPUs. It would be quite interesting to see how other CPUs fare in the benchmark. Also similar comparison with SpecInt would be interesting to see. I wouldn't be surprised to see a few benchmarks bottlenecking there too.

Aguia said...

It is probably just my fault of not doing enough research but what exactly is so special about those chipsets?

They normally work. Right from the beginning.

Capable of playing games.

The on paper specifications match the hardware/software specifications.


This will probably ansvered by previous question but what exactly does Intel need to catch up with?

Aguia said...

I'm qurious, what benefit does HT3 have against HT1 on desktop?

Ho ho it was you who wrote this right?

First, no it does not. Two 8800Ultras beat it just a few days after XT's did the record.

Let me list this for you:
AGP 8x 17.066 Gbit/s 2.133 GB/s
PCI Express (x16 link)32 Gbit/s 4 GB/s
AGP 8x 64-bit 34.133 Gbit/s 4.266 GB/s
HT (800 MHz, 16-pair) 51.2 Gbit/s 6.4 GB/s
HT (1 GHz, 16-pair) 64 Gbit/s 8 GB/s
PCI Express (x32 link)[7] 64 Gbit/s 8 GB/s
PCI Express 2.0 (x16 link) 64 Gbit/s 8 GB/s
PCI Express 2.0 (x32 link) 128 Gbit/s 16 GB/s
HT (2.8 GHz, 32-pair) 179.2 Gbit/s 22.4 GB/s

Intel bus speed/Bandwidth:
800Mhz 6,5GB/s
1066Mhz 8,6GB/s
1333Mhz 10,8GB/s

Intel shares the bus with memory data and PCIe data, AMD only PCIe.

Ho Ho said...

As can be seen from current CPUs and chipsets Intel shared FSB is enough to feed even the fastest GPUs on the market. This is no miracle as PCIe has considerably less throughput and in multi GPU solutions there is a special bridge between the cards.

Do you have anything besides theoretical peak bandwidth numbers to support what you are trying to say? E.g some benchmarks comparing older 800MHz HT vs 1GHz HT. If there is some benefit it should show up there, wouldn't it?

Aguia said...

It may not be the fastest performing CPU, but it's for sure the best platform and best architecture, period.

AMD's Barcelona architecture

Scientia from AMDZone said...

ho ho

"That 3GHz box was sitting in inaccessible place (doing nothing but running OS?)"

It ran the graphic demos.

"and nobody outside AMD could touch it. Why is that?"

For two reasons: You can't do that in that kind of venue; you need smaller groups. Secondly, AMD has no reason to announce benchmarks until the chips are officially released.

"Also as you probably saw from the slides AMD doesn't expect to reach 2.6GHz on quadcores before Q2 next year."

You would have to tell me what slide you are talking about. The Roadmap slides have no clock speeds listed. The only other slide I'm aware of only covers to Q4 of this year and only says "2.3Ghz and above".

"All the differences you pointed out should give slight but most likely neglible speed decrease for the slower system. What do you think how much could have the results changed if both systems were configured exactly the same?"

Yes, it is puzzling. Network support shouldn't make that much difference either.

"Has anyone got any idea what wattage will the initial CPUs be?"

I believe the 2.0Ghz chip is 95 watts and below that is 68 watts.

" 120W at 2.8 and 3GHz. As I said VMWare slide backs up this theory."

True but I imagine the VMware testing was done on the previous stepping. Slide 47 shows what the chips will be. It says:

HE - 1.9Ghz at launch (68 watts)
Standard 2.0Ghz at launch (95 watts)
SE - 2.3Ghz and up Q4 (125 watts)

So at launch:

95 watts - 2.0Ghz
68 watts - 1.9Ghz and below

Q4:

125 watts - 2.3Ghz
95 watts - 2.1 - 2.2Ghz
68 watts - 2.0Ghz and below

"Another thing I'd like to know is where did you got that 15% IPC increase compared to K8 in SpecFP."

As I said before, you can only compare the total SPECfp number at the same clock; you can't compare the subtests. The score is 15% higher; I don't think that means K10 only has 15% more IPC.

Scientia from AMDZone said...

ho ho

"You do remember that Intel doesn't ramp before it has squeezed out almost everything from the node"

That isn't true this time. Intel's G0 stepping looks a lot better than the F2 or F3 steppings yet Intel will move to 45nm before the G0 stepping is really utilized.

" and AMD releases a bit earlier and does continous improvements on the transistors?"

True again but AMD also tends to develop the previous process further than Intel. AMD's 90nm is without doubt better than Intel had. It also appears that AMD will surpass Intel's current level with 65nm and continue where Intel might have gone with the G0 stepping.

"You probably meant to say"

All advantage in terms of battery life and performance while using batteries. And, brand recognition is not what you think it is. One consumer study showed that most consumers cannot remember which brands are Intel's. Intel complained that its competitors picked similar sounding names to purposefully confuse consumers. Xeon, Athlon, Duron, Celeron, Sempron, Opteron: Yes, I can see how this would be confusing.

You have to admit that Core 2 Duo is an absurd name for branding. They could probably get some pointers from Apple who still use a variation of Mac.

"We can only assume that Phenom is faster than K8 and Penryn is faster than Core2"

And Shanghai is faster than Phenom.

"Power is also quite nice but probably only competes with Itanium"

No. x86 server chips from both AMD and Intel are slowly consuming Power's market. Itanium is increasing in volume but not gaining in share. Power is losing share.

"How high were the margins in Q2 and how high will they rise in Q3? Above mid-40's? Somehow I doubt they recover that fast."

If we exclude the inventory write-off AMD's gross margin in Q2 would be 35%. It should be 37% next quarter. It could be about 40% in Q4. However, he was talking in general, not just the near future.

"I take you don't think that continous price cuts throughout Q2 and beginning of Q3 doesn't affect AMD margins a lot."

Not to that extent, no. Think of it this way. In Q1 we had a huge difference in gross margin. However, in Q2 Intel's margins dropped while AMD's rose. We should see the two coming closer together over the next several quarters. What we will see is that without 200mm to reduce AMD's margins and without having delays because of having to build a new 300mm FAB AMD's margins will slowly gain on Intel. In the future Intel will not be able to drop its prices without hurting its own margins.

"Intel orders are huge part of the market share it will never get back though."

It doesn't need to. As I mentioned, the 690G chipset is becoming very popular and this should continue. AMD's discrete grapics chips should also be back on track in Q4. Then in 2008 AMD's 780 mobile chipset should become the dominant mobile chipset for AMD platforms.

"can copper wire based HT with lots of wires going from CPU to motherboard can compete with silicon optics?"

When do you suppose this will actually be available?

"It is probably just my fault of not doing enough research but what exactly is so special about those chipsets?"

My understanding is that it's the graphics capabilities. It seems to be enough to make 690G more popular than nVidias' chipset. Wouldn't you agree that 35 motherboards using it is popular?

nVidia and AMD currently have the advantage because they produce discrete graphics chips. This means that as a given GPU design becomes dated it then becomes the core of the next chipset. Intel doesn't have these discrete graphic chip designs so it doesn't get this trickle down benefit.

"What exactly tells us that AMD can gain any marketshare with its new platform?"

They have been gaining since they released Turion.

" Sure, it does have better power managment than before"

Griffin is a K8 that has been optimized completely for power consumption. It has a split power plane for the memory controller and can also reduce power on HT. It can reduce clock on each core separately and the memory controller has been optimized for just two SODIMMs. It matches Intel's Tubo Memory with HyperFlash. However, the RS780 chipset includes the ability to switch between integrated and discrete graphics when on batteries or plugged into the wall.

I seriously doubt that quad core is going to be in big demand on mobile. Secondly, Intel slipped in power consumption when it introduced Merom and has been slipping again. Intel will need to get its power consumption back down or it will be behind AMD.

This is one of the fundamental questions. AMD is using a different architecture for Griffin while Intel is using the same one. It will have to be seen whether Intel can maintain the low power draw with the same architecture it uses for the desktop. Intel could easily have an advantage in performance when plugged into the wall but that won't matter if they can't match AMD's power draw or performance/watt when on batteries.

Ho Ho said...

aguia
"It may not be the fastest performing CPU, but it's for sure the best platform and best architecture, period."

Best for what? Everything or only some special work cases?


According to SpecFP_rate K8 seems to be better platform than Xeon also but for some reason it still massively lost marketshare. This shouldn't happen if you have the overall best platform on market.


scientia
"It ran the graphic demos."

Yes, I just found it out myself just half an hour ago. I wonder why it was reported otherwise before.


"You would have to tell me what slide you are talking about."

I saw those slides a week or two ago, can't remember where they originated but here seems to be one of them.


"Yes, it is puzzling. Network support shouldn't make that much difference either."

Fact is it can't. I don't even understand why did you brought it up as it is the last thing that could affect performance.

Also as I said if there is speed decrease then only for the slower system. Did you miss that part when you offered those things as theories of why there is this kind of difference in results?


"True but I imagine the VMware testing was done on the previous stepping."

Probably so. I thought that it sounds a bit too bad if AMD really would have those 2GHz ones at 120W+. It would have made it difficult to believe they didn't need to go through lots of CPUs to find the ones that would work at 3GHz for the tests.

Then again they will only have 2.3GHz out before the year end at 125W, nearly half a year for 300MHz increase combined with 30W TDP increase. Makes one wonder what it takes to get mass production of 3GHz with half-decent thermals.


"As I said before, you can only compare the total SPECfp number at the same clock; you can't compare the subtests."

So basically that SpecFP_rate result is nearly meaningless as you can't compare it to anything and expect to get any reliable result out of it.


" The score is 15% higher; I don't think that means K10 only has 15% more IPC."

It was ~70% faster with twice the number of cores. Had it recieved identical results in memory bottlenecked benchmarks as did K8 it should have got significantly higher results as there were plenty of benchmarks where bandwidth shouldn't have been a problem. As aguia link said AMD expects to have 50% more memory bandwidth availiable for K10 compared to K8 and that alone should give it quite a boost. Something tells me that things mithg not be all that good afterall.


The memory system improvements should give AMD at least as much as does 1.33GHz FSB to intel. Even still basically identical cores on already memory bottlencecked Core2 scale nearly 69% from 2GHz dualcore to 2GHz quadcore then how could much better K10 scale just as bad? It should do much better, especially considering all the improvements in core.

I just added the quadcore to the charts:

averages
scaling

abinstein said...

Ho Ho -
"If clock speed would be the only thing affecting the benchmark then we should see roughly 22.5% speed difference in every sub-benchmark."

All tests have different mixes of branches, memory loads, ILP, etc. How can you expect they scale the same? Furthermore, the sub-tests are the kernels of some specific application. Thus any sub-test alone is not meant to nor supposed to represent general performance.

"K10 with 8 2GHz cores has a score of 69.5. With half the number of K8 cores AMD scored 40.7. Increase core count by 100%, gain 71%. Kind of small gain ..."

The 2.0GHz Barcelona SPECfp is actually about 78 from this video. AMD probably made up that 69.5 by estimation. 78 is 1.9x of 41 achieved by dual-core Opteron. Note earlier the 92 @2.6GHz has ~18% higher score with 30% faster clock.

All your other "analyses" become moot, don't they?

"Unfortunately they don't tell which one exactly but if it is DDR2 one then keep in mind it has TDP of at least 119.2W, unless they compare it with 1P one at 125W."

AFAIK, AMD do not measure processor, but system power. It's possible that the dual-core K8 with the same memory & peripherals takes 60W more (30W per socket) than the quad-core K10; in terms of system wattage, it's probably just 20% more. Now, 1.79/1.2 = 1.49, so in other words, 2GHz quad-cre K10 is 49% better than 3GHz dual-core K8, and consumes 20% less power.

"That 3GHz box was sitting in inaccessible place (doing nothing but running OS?) and nobody outside AMD could touch it."

Now read this article from someone actually attended the event.

It seems to me making FUDs and speaking of stuff that you don't know has become a really bad habit of yours.

Scientia from AMDZone said...

ho ho

What advantage is HT 3.0 on a single socket desktop system?

HT 3.0 will probably be used to connect the two dies when Shanghai goes to eight cores. There would be a bottleneck with just HT 1.0.

Faster than HT 1.0 is also important if you are using integrated graphics with no frame buffer.

Scientia from AMDZone said...

ho ho

"I saw those slides a week or two ago, can't remember where they originated but here seems to be one of them."

Yes, those are from February. At the time, Intel's fastest quad core in 2007 was projected to be 2.66Ghz so 2.5Ghz would have been sufficient. And, the 2008 numbers were projected more than a year in the future. AMD's recent 3.0Ghz demo shows that it is ahead of this schedule. I would guess Q1 for regular 3.0Ghz quad cores.

"I don't even understand why did you brought it up as it is the last thing that could affect performance."

I didn't bring it up; you did.

Ho Ho said...

scientia
"Intel's G0 stepping looks a lot better than the F2 or F3 steppings yet Intel will move to 45nm before the G0 stepping is really utilized."

Define "really utilized". Intel will be producing 65nm chips for at least until the end of 2008, it has plenty of time to utilize them. You could just the same say that AMD has little reason to improve its K8 as it will "soon" be replaced with K10 but it'll still have 6400+ on market.


"True again but AMD also tends to develop the previous process further than Intel"

I'm curious, what metrics do you use to say that AMD develops their process further than Intel?


"AMD's 90nm is without doubt better than Intel had."

Pentium M was rather nice at 90nm.

"All advantage in terms of battery life and performance while using batteries."

Battery life, perhaps, but most certainly not performance. I highly doubt AMD can clock its 65nk K8-derivate nearly as high as Intel its 45nm ones and Intel should have much better IPC. It is not K10 we are talking, after all.


"One consumer study showed that most consumers cannot remember which brands are Intel's."

People doesn't need to know what CPU they have in their PC, they just remember the brand name "Centrino".


"You have to admit that Core 2 Duo is an absurd name for branding."

I personally couldn't care less what names are used. I only care about nubmers and sometimes codenames as it is usually easier to remember them.


"And Shanghai is faster than Phenom."

... and Nehalem is faster than Penryn :)


"Itanium is increasing in volume but not gaining in share. Power is losing share."

We shall see what the new Power6 does to the market. It surely is an impressive CPU. Though it's kind of funny that it is only slightly better than nearly 3x lower clocked Itanium in SpecFP :)


"If we exclude the inventory write-off AMD's gross margin in Q2 would be 35%. It should be 37% next quarter. It could be about 40% in Q4. However, he was talking in general, not just the near future."

Assuming AMD won't have any more write-offs and continues to increase its margins it should reach upper-half of 40% at the end of Q2 next year or nearly a year from now. I guess it depends on how one defines "near future". Also I'm quite sure AMD will have to lower its K8 prices again this year. If not sooner then after Penryn comes and pushes current Core2 prices down.


"It doesn't need to. As I mentioned, the 690G chipset is becoming very popular and this should continue"

As Intel does have majority of the marketshare and AMD CPUs are powered by NV chipsets too then loosing Intel is quite big deal.


"AMD's discrete grapics chips should also be back on track in Q4"

I'd like to see that but I wouldn't bet my money on it. With FX series NV was similar to Intel and Netburst. After the FX it got its act together and has gained massive marketshare since then. Differently from Intel it has managed to do it with trailing technologies (low-k with R420, R520 first at 90nm, R600 at 80).
By the years end we will probably have dual-GPU RV670 fighting against 65nm G92. NVidia has said it will have up to twice the pefrormance of old GTX and it isn't totally unbelieveable. I have hard time to believe that AMD can fight against that amount of power before R700 comes some time next year (Q2, Q3?). We should also see 65nm low-end refreshes. Certainly difficult things to counter, especially considering AMD is currently fighting against NV 80nm with 65nm chips.


"When do you suppose this will actually be available?"

My best guess would be around 2009/2010 if everything goes fine. If not then perhaps a year later.


" Wouldn't you agree that 35 motherboards using it is popular?"

Well, number of different boards is probably not as important as number of total boards sold. I guess we don't have that number yet. Before that I wouldn't make too much predictions.
Btw, do you know how many 965 based motherboards are out there?

Fun but pretty much useless fact: there are four 690 based motherboards listed on estonian pricewatch site. Against those four we have 25 motherboars with 965 chipset. Seems as 965 is quite popular also.


"They have been gaining since they released Turion."

Of cource they have. Going from zero to something is a massive gain. My point was that getting a new platform on the market won't probably accelerate the process too much.


"I seriously doubt that quad core is going to be in big demand on mobile."

It will depend on what can Intel do with the thermals.


"AMD is using a different architecture for Griffin while Intel is using the same one."

This is something I wanted to talk about. Intel has two teams working on x86 in tic-toc mode pumping out new cores and refreshes every two years. AMD Pipe seems to be something similar. Will it have additional two teams working on its mobile CPUs?


"Intel could easily have an advantage in performance when plugged into the wall but that"

This might very well become true. Though it should be known that mostly people use their laptops just the same way as desktops: connected with wires. I've seen a lot of people who want to change their PCs and get laptops to replace them. Not because they suddenly need to be mobile but just because they take less room and are considerably more simple to carry around to e.g work from home. Of course for some people working on batteries is the most imortant metric but you can be sure there are more people who only use laptops as portable PCs.

sharikouisallwaysright said...

http://www.heise.de/newsticker/meldung/93607

They are telling, that some of the trouble 2006 to 2007 comes from Dell has rejected to take as many CPUs as they initially wanted to take from AMD and AMD had very big trouble to sell the chips otherwise.

Aguia said...

You have to admit that Core 2 Duo is an absurd name for branding.

Good point. I even think that after Intel named their processors, Core 2 Duo
What is holding VIA or AMD to release their processor by similar name.

AMD Athlon 64 X2 Dual-Core Processor
to
AMD Core X2 Dual
AMD X2 Core Dual
AMD X2 Dual Core
AMD Dual X2 Core
AMD Dual Core X2

Not bad hum?


Best for what? Everything or only some special work cases?

I see you don’t seam to know what best means. I wonder when Intel releases Nehalem with technology that AMD already as in Opteron since April 2003 and Barcelona since September 2007, I would like to see what you will say, I hope you won't announce them as being new to the market or as being the ultimate new techology.

Best <> Fastest!!!
Its not
Best == Fastest!!!

For example to me Ati 2600 series is better than Nvidia 8600. Because of small detail factors not performance. In fact I post the 4 reasons why Ati 2600 series is superior if you would like to know.

HT 3.0 will probably be used to connect the two dies when Shanghai goes to eight cores. There would be a bottleneck with just HT 1.0.

Wrong. It’s because of PCIe 2.0.
If that’s the case Barcelona was the First to receive HT 3.0 and not the desktop system. HT 1.0 would limit available bandwidth for the new PCIe 2.0 GPU’s, used or not.
Your answer only gives ho ho point of view as being the correct one.

Aguia said...

I would guess Q1 for regular 3.0Ghz quad cores.

I have already showed that Intel demo Conroe at 3.5Ghz and Kentsfield at 3.33Ghz and none of this CPUs still exists today.
Intel and AMD high clocked CPUs, keep dreaming!

enumae said...

Abinstein
The 2.0GHz Barcelona SPECfp is actually about 78 from this video. AMD probably made up that 69.5 by estimation.

Could that be the average of one of the 17 test in SPECfp?

It seems strange that AMD would not have used those results and compared them to Intel's best published results, isn't it?

If they can achieve a score of 78, that would be 44% faster than Intel's best published score, it would also be a real score, not an "estimated result", as per the PDF from the analyst day.

abinstein said...

enumae -
"It seems strange that AMD would not have used those results and compared them to Intel's best published results, isn't it?"

If you've watched the video, it's obvious why AMD didn't use the higher number. In the side-by-side demo, the 2.33GHz Xeon reaches a score 60, higher than any published score. Apparently AMD's engineering systems (both hardware & software) perform better than the commercially available ones.

People will cry foul if AMD compared Barcelona's 78 (actually 77.5 rounded up) with Xeon's published 54. What AMD does is simply scale back Barcelona's score the same amount as the Xeon (exactly 10%).

abinstein said...

aguia -
"If that’s the case Barcelona was the First to receive HT 3.0 and not the desktop system. HT 1.0 would limit available bandwidth for the new PCIe 2.0 GPU’s"

Quite accurate. It is very unlikely that AMD will make oct-core processors with MCM on the desktop space. The main reason is MCM with HT will inevitably result in a NUMA system, which most desktop applications don't optimize for. Just take a look at QFX and one see how improper it is for desktop (enthusiasts).

BTW, this is not the case for the workstation or server space.

lex said...

Give you credit you write like you know architecture and design but have can't fake like you have a clue to how the silicon side of the business works. That is the one competitive advantage that companies like TSMC, Samsung, and INTEL leverage that create a huge competitive moat around their business. It is also why some companies that once had great silicon technologies have fallen off to the side; Motorola, TI, and IBM to name a few. I can see that most of the analysts are easily impressed by fancy PowerPoint showing impressive product plans and don't appreciate the complexity of the silicon manufacturing that has to happen as well as the technology elements and work involved in getting them and you are clearly in that camp too! Some think you just go to Applied, ASML and put those tools in the fab and viola you got high performance, high yielding silicon just like that. But unfortunately just like football you can get those athletes but if you don't have a solid program and do your homework you won't have a great winning football program. As such AMD will struggle mightily at 45nm and 32nm. Neither will give them equality to the INTEL but their disadvantage is actually growing. They haven't put the necessary leg work in for the past few years and are fooling everyone and themselves if they think they can pull it off with the spending limits they have.

I'll just address a few of your comments.


"I'm not sure what you are trying to say. FAB 36 will probably be 7-9 months behind in 45nm. We don't know yet if this will be slower"

WRONG, if 90 and 65nm are any indication AMD/IBM process has been inferior to INTEL offerings. So they are both late and also inferior. At 45nm that performance gap will be even larger with the fact the 1st 45nm from IBM won't have HighK/Metal gate. At these dimensions not having the ability to scale gate oxide to control transistors at the expense of gate leakage will really penalize AMD's offering in 2008. My estimate that AMD is going to be stuck using traditional oxide it will be a penalty them by either 2x in gate leakage or 20 percent in raw performance. That is a lot to expect your design to makeup for. Again my thesis is unless Nehalem is a bust 2nd half 2008 AMD will be like the 1st half of 2007 bleak and red as AMD has to cut prices to slot their slower products competitively.

"AMD already has a 45nm immersion scanner in operation. Test wafers are running now and we should see a Shanghai tape out soon."

How is that impressive that you need to make a significant lithography change? That sounds very risky to me actually. INTEL seems to have gotten by with tried and true dry lithography. AMD is going to be going first with wet litho, new resist, new tools and I’m sure lots of surprises yet to come. I believe if AMD does ramp as they claim they will they will be the first to go to production with immersion I can only think back to all the 1st in silicon and the pain that was born by the innovator; LowK, DUV, 193nm etc. etc. I have no confidence that AMD has the resources, deep pockets, track record as a process leader ( oh.. I'm joking here! ) nor the most important thing TIME to do this. Oh, lastly if they are running only test wafers now they are in big trouble. The full capability of any new technology always is learned ramping some serious pattern sensitive debug vehicle. If all they got are test wafer running Shanghai will get shanghaied for sure and show up very very late or not at all.

"Not exactly. There are both advantages and disadvantages to having more volume. AMD should have a much easier time ramping FAB 38 than it did ramping FAB 36."

There is NOTHING to be gained by a small fab. The technology problems are the same whether you are running a few thousand wafers a week with a couple tools of each type or much larger scale with even more tools. Nothing is worst then having a process issue where you can't compare your output from 3 or 4 different tools to figure out what is broken. The overhead is also fixed, the bigger your fab the larger the volume, revenue you can spread the huge fixed costs of engineering, facilities etc. etc. A SMALL fab is absolutely less efficient then a larger one by any metric.
Thus a small fab is the worst thing for AMD, they need a Mega fab to maximize resources and distribute the fix cost of the factory.

"Your mistake is in seeing these as continuous expenses rather than temporary expenses"

Sorry is R&D and future manufacturing capacity investment a temporary expense. You are a funny man. When do 32nm R&D and manufacturing costs start? Isn't AMD trying to catch up? I would expect to see some serious expenditure in 2008 for 32nm if they are to match INTEL. That’s a heavy burden to invest in 45nm this year and 45nm AND 32nm next year. Or are you assuming AMD will continue to lag INTEL by a year+ and expect design to make up that gap. That is where Intel’s tick tock really will crush AMDs design/process conversion cadence.

"AMD's only competitor in processor manufacturing is Intel." DUDE if there are only two players there is FIRST and their is LAST. AMD is in the bottom half, last I checked by every metric AMD is LAST in the cellar, LOL. Sorry, don't tell me about YOY growth when you are starting from a miniscule number AMD numbers always look distorted. The bottom line is AMD MS encroachment have been stopped and if anything its all going back to INTEL across most segments. Lastly it would be okay if AMD was LAST out of two and still MADE money. Can you please aggregate AMD's cumulative profit/loss for the last 1, 5, and 10 years and compare them to INTEL. You seem to like numbers. Tell what those numbers tell you about AMD and whether those fancy roadmaps are going to appear or not.

Sorry unless they find a whiteknight AMD is finished, they don't have the cash flow to fund the technology they need to compete. There is nothing to argue, wait and see what happens end of next year when 45nm is suppose to happen for AMD .

enumae said...

Abinstein

Sorry, but you avoided the main question so I have to ask it again...

She says "were running some of the SPEC benchmark", and "were running it over and over again".

Looking on SPEC.org, it could take a couple of days to run SPECfp, int and cpu, so, could they be running just one of the test inside of SPECfp?

abinstein said...

"Sorry, but you avoided the main question so I have to ask it again..."

Sorry, but I'm not a mind reader, and if you have some "main" question to ask me, mark them specifically so next time, or I'll just ignore (not just avoid) whatever I think they are stupid.

As to your "main" question, yes it is possible that the demo is just a sub-test. If you compare the score, it's likely the 434.zeusmp - 58.0, or maybe 436.cactusADM - 66.1. Both scores are taken from this source.

Interestingly, dual-core Opteron performs both tests somewhat slower than dual-core Xeon. For 434.zeusmp, K8 is only 92% the speed of Core 2; for 436.cactusADM, K8 is only 84% the speed of Core 2.

In other words, AMD is not Intel, and AMD does not play dirty tricks. It is actually picking the fairest floating point metric, slightly conservative in favor of the competition, to demostrate the improvement of K10 over both Xeon and K8.

yomamafor2 -
"AMD's fastest beats Intel's slowest!! That's the big turn around of the century!!"

The right way to interpret the results is this: for SPECfp, dual-core K8 beats Core 2 Duo, and quad-core K10 beats Core 2 Quad.

BTW, 2.33GHz is not Intel's slowest. According to AMD, as of today, it is Intel's fastest processor under 80W TDP.

abinstein said...

yomamafor2 -

Some update from my previous comment -

1) For SPECfp, dual-core K8 beats Core 2 Duo, and quad-core K10 beats Core 2 Quad even more.

2) According to Intel, 2.33GHz is Intel's fastest processor under 80W TDP today.

abinstein said...

lex -
"It is also why some companies that once had great silicon technologies have fallen off to the side; Motorola, TI, and IBM to name a few."

What you're talking about here is the amount of capital a company puts into semiconductor production. The more capital you have, the better capacity and faster ramp you can get, and the lower per-unit cost your products have. However, capital or capacity is not the only thing that matters; there are also technology and management. AMD do not have the money to compete with Intel alone, but it is cooperating with IBM, Chartered, and TSMC to get to the next process node.

You are also terribly wrong to compare Intel to Samsung and TSMC. Intel makes CPU and chipsets for itself, and that's pretty much it, while TSMC is a foundry and makes chips for everybody. Samsung makes a lot of memory, but their foundry business is just beginning in 2006. Samsung isn't even one of top 10 foundries.

You are also terribly wrong in suggesting TI or IBM "fallen out of the side." TI has great SoC capability down to 45nm; while SoC is (so far) not helpful to CPU, it is great for pretty much everything else. IBM still holds the technology crown of foundries, not even TSMC can compete with it. Intel just completely lost its face in not getting 4GHz before IBM.

"... just like football you can get those athletes but if you don't have a solid program and do your homework you won't have a great winning football program."

Semiconductor manufacturing is not just like football. I don't know where you get that idea, but your analogy is ridiculous.

"As such AMD will struggle mightily at 45nm and 32nm. Neither will give them equality to the INTEL but their disadvantage is actually growing. They haven't put the necessary leg work in for the past few years and are fooling everyone and themselves if they think they can pull it off with the spending limits they have."

You really like babbling away, don't you? Here are some head-clearing facts and questions for you to ponder upon:
1) Who will not struggle mightily at 32nm? Intel?
2) How so you know AMD didn't put necessary leg work?
3) AMD doesn't need to spend as much as Intel to get to the next node. It has been so since the days of 130nm.

"INTEL seems to have gotten by with tried and true dry lithography. AMD is going to be going first with wet litho, new resist, new tools and I’m sure lots of surprises yet to come."

I have to say: you write like you know architecture and design but can't fake like you have a clue - does this seem familiar to you?

So AMD is the first in mass wet production. Isn't Intel the first in mass double patterning? And are you sure or do you have any idea what "new tools" or materials Intel uses in its 45nm?

"If all they got are test wafer running Shanghai will get shanghaied for sure and show up very very late or not at all."

You can't even understand the difference between engineering test wafer and production test wafer, and how do you think you are qualified to make prediction?

"DUDE if there are only two players there is FIRST and their is LAST. AMD is in the bottom half, last I checked by every metric AMD is LAST in the cellar, LOL."

This is probably the only thing that you said with any sense in it. Intel made money before, and AMD lost it. But if you want to talk about business, you should look toward the future. The worst thing one can do is to predict the future based on the past, and unfortunately that is precisely what you are doing here.

If you don't know about some technology, refrain from speaking or predicting it.

enumae said...

Abinstein
Sorry, but I'm not a mind reader...

"Could that be the average of one of the 17 test in SPECfp?"

Thats from my post in which you replied, and all my other statements were relative to that question.

Please drop the sarcasm (as I have asked before), the question was clear.

As to your "main" question, yes it is possible that the demo is just a sub-test.

Thank you for answering my question.

In other words, AMD is not Intel, and AMD does not play dirty tricks.

If it was a subtest, and AMD had you, the journalist, and other people who have watched the video, believing it was a full SPECfp test, you don't feel it is misleading????

I consider you much smarter than I am, but that is misleading.

I don't know why we always get off on the wrong foot...L8TR

Unknown said...

enumae
Looking on SPEC.org, it could take a couple of days to run SPECfp, int and cpu, so, could they be running just one of the test inside of SPECfp?

I think AMD focused on running SPECfp, because that's where AMD will definitely shine and stand out.

In other words, they purposefully ignored other tests, so the public will focus on "30% better performance per watt per dollar per PR stunt".

abinstein said...

enumae -
"Thats from my post in which you replied, and all my other statements were relative to that question."

Sorry, but I fail to see the relation. SPECfp or sub-test, what does it change? I showed you that if the test was a SPECfp, it is only more in Intel/Xeon's favor. In other words, AMD's claim that Barcelona is superior than Xeon would be even more true if the test was one of the two possible SPECfp sub-tests.

Yes the question was clearly there but I deemed it not worthy of a reply.

"If it was a subtest, and AMD had you, the journalist, and other people who have watched the video, believing it was a full SPECfp test, you don't feel it is misleading????"

Please, just read my comment and tell me how does this sub-test mislead any conclusion? There are only two SPECfp tests have score around 60 on 2.33GHz Clovertown; both favors Xeon more than K8. There is a good reason for AMD to demo this single benchmark rather than the full SPECfp, which would run not even once in several hours - and you've said it yourself.

If AMD were Intel and wanted to play dirty, it could've found at least 5 other SPECfp tests that make K10 look much, much better than Xeon. That would be misleading.

At the end of the day you just need to get one conclusion home: Barcelona is having much better fp than both Clovertown and K8.

Ho Ho said...

abinstein
"All tests have different mixes of branches, memory loads, ILP, etc. How can you expect they scale the same?"

First you missed the part where I said "If clock speed would be the only thing affecting the benchmark". Secondly the CPUs had only difference in CPU clock speed.

If you still don't understand then the purpouse of the comparisons I have made is to show that there are huge differences between the subtests in Spec benchmarks and average score scales more with memory bandwidth than with CPU speed making any kind of IPC calculations pretty much impossible.


"The 2.0GHz Barcelona SPECfp is actually about 78 from this video. AMD probably made up that 69.5 by estimation."

First, how big time difference was there between the analyst day and the day that video was made? Do you really think that AMD couldn't get any real SpecFP numbers before analyst day and it had to yet again simulate them?

Of cource as was later said, this is not overall score. If you look at the closeup screen you'll see that the wholse screen is filled with numbers around 78. Spec FP only runs every test three times. My guess is that they simply showed one benchmark that was in the middle of execution and those numbers were just some intermediate results from the specific benchmark. Not the score of the benchmark but just some results. For one thing I couldn't see the execution time for benchmarks. There were only three columns: iteration, iteration time and running average.


"All your other "analyses" become moot, don't they?"

Until we get any kind of confirmation they do not. Anyone wants to make a guess if, what and when will we see anything new here?


"AFAIK, AMD do not measure processor, but system power"

When they label CPU as 125W TDP then they mean the CPU, not CPU+half the system. You can't make such a comparison as there are different kinds of systems out there each with different thermals. Now that does the rest of your analysis moot :)


scientia
"HT 3.0 will probably be used to connect the two dies when Shanghai goes to eight cores. There would be a bottleneck with just HT 1.0.

Yes, for that it could help. Though I was thinking more like what is the use of having faster connection between CPU and northbridge when Intel slow and shared FSB is good enough.


"Faster than HT 1.0 is also important if you are using integrated graphics with no frame buffer."

It could help for sure but first we would need to get significantly faster memory to make use of increased bandwidth.


"Yes, those are from February. At the time, Intel's fastest quad core in 2007 was projected to be 2.66Ghz so 2.5Ghz would have been sufficient"

In february they also meant to release K10 at much higher frequencies. That was before they knew they would have problems with that. Wouldn't those problems push back the schedule even more than what was predicted in February?


"AMD's recent 3.0Ghz demo shows that it is ahead of this schedule. I would guess Q1 for regular 3.0Ghz quad cores."

I highly doubt that. I'm still quite sure we won't be seeing quads nearly that fast before H2 next year.


"I didn't bring it up; you did."

I showed the benchmarks, you were telling they use different setups. I told these differences are so small they don't have an effect. Now you say I brought the subject up?



aguia
"I see you don’t seam to know what best means."

I know that Core2 is best for SSE heavy job and K8 is best for memory intensive ones. K10 will make the difference smaller but pure clock speed difference and lack of SSE4 still means that Core2 will lead the high end with SSE performance. There is no overall best CPU that is best in everything.


"For example to me Ati 2600 series is better than Nvidia 8600. Because of small detail factors not performance. In fact I post the 4 reasons why Ati 2600 series is superior if you would like to know."

So if you ignore the speed as one of the metrics you can list four reasons why it might be better. Ok, please do then. Will Linux support also be discarded?


"HT 1.0 would limit available bandwidth for the new PCIe 2.0 GPU’s, used or not."

Then please explain me why doesn't even the 1066MHz FSB limit GPU performance on Intel machines.
Btw, main reason for PCIe2 is to increase power delivery to the slot. Even today you must be looking really hard to notice speed differences between 8x and 16x PCIe.



abinstein
"Apparently AMD's engineering systems (both hardware & software) perform better than the commercially available ones."

It is quite interesting they get around 10% better performance with Intel than anyone else before them ever has. If they didn't use any tricks then it is kind of difficult to not repeat their results. For some reason none of the twelve submitted results even get near it. Interesting, isn't it? Can you list a few things others haven't used that could increse Intel score by around 10%?

Of cource alternative is that you discard it as overall result.


"People will cry foul if AMD compared Barcelona's 78 (actually 77.5 rounded up) with Xeon's published 54. What AMD does is simply scale back Barcelona's score the same amount as the Xeon (exactly 10%)."

An interesting theory. Only thing missing now are facts that could prove it.


" Intel just completely lost its face in not getting 4GHz before IBM."

Sure, clock speeds are the most important thing in the world. Well, they are if your competitor has nearly the same IPC but it doesn't when CPU clocked nearly three times slower manages to do the job with same time period. Only thing good about Power6is its massive L3 cache (look, a CPU built around cache! Must be a dirty trick to get benchmark wins! [/sarcasm]) and its massive connection speed.


"1) Who will not struggle mightily at 32nm? Intel?"

Ask again in around two to three years when we have some data. Who would have thought that AMD has "slight" clocking problems with K10 just half a year ago when Inquirer reported massive clock speeds for it?

abinstein said...

Ho Ho -
"My guess is that they simply showed one benchmark that was in the middle of execution and those numbers were just some intermediate results from the specific benchmark"

Yes, and please follow my previous comment and see what "specific benchmark" it could be. You will see it could be one of two benchmarks which Clovertown had clear advantage over K8 at the same clock rate.

K10 is simply superior to both Clovertown and K8.

"It is quite interesting they get around 10% better performance with Intel than anyone else before them ever has."

No, you are wrong. AMD actually had its Xeon working at either 1% faster (for 434.zeusmp) or 10% slower (for 436.cactusADM) than the best published score.

"Only thing good about Power6is its massive L3 cache"

Have you any idea about processor microarchitecture? Your comment here is almost shamelessly Intel biased.

"Ask again in around two to three years when we have some data. Who would have thought that AMD has "slight" clocking problems with K10 just half a year ago when Inquirer reported massive clock speeds for it?"

First, lex, who I was replying to, did not wait until 2009 to make his pompous statement. What he seems not understanding is that everyone on the planet working on semiconductor today is scratching his head on how to get 32nm and below work.

Second, the "slight" problem apparently is not on clock speed, but on yield. And as you said, the problem is really slight compared to those people are facing for 32nm.

abinstein said...

Some old comments that I missed....

enumae -
"AMD 2P = 52.5
Intel 2P = 43.1
AMD has an 20%+/- advantage."


You are comparing 3.0GHz K8 with 2.66GHz Core 2. At the same clock, dual-socket K8 X2 is just about 10% faster than dual-socket Core 2 Duo in SPECfp.

Ho Ho -
"Also, abinstein, do you still think that K10 won't be bottlenecked in SpecFP as you so politely said to me in sharikous blog?"

Performance scaling is not just core clock or bandwidth, but other things (such as memory latency, branches, forwarding, front-end, etc.) I am amazed that you would compare the scalability across different sub-tests and claim the only effect is memory bandwidth simply because 1) they run on the same processor family, 2) you don't know better otherwise. No no no, different sub-tests have different memory access, branching, cache miss pattern, and thus are scaled by clock rate quite differently.

If you had actually compared the sub-tests scaling closer and more properly, you'd have seen there are SPECfp tests that Core 2 do extremely bad while K8 scales very well (410, 481, 482). These alone can prove your "bandwidth limit" theory wrong.

"Though I wonder why didn't you make any more comments on what I said. Do you agree or disagree with what I said about scaling and did I got the numbers correct?"

Have you thought of it's probably because I'm tired of replying the same thing to someone who makes so many errors but just can't learn and don't correct himself.

Aguia said...

So if you ignore the speed as one of the metrics you can list four reasons why it might be better. Ok, please do then. Will Linux support also be discarded?
Then please explain me why doesn't even the 1066MHz FSB limit GPU performance on Intel machines.


Your problem Ho ho is that you measure performance above anything. There are other more important factors than performance.
If performance was the most valued metric, Cyrix and AMD, was already out of business many years ago.
Well Wii beats X360 + PS3 in sales, but is the worst of all in performance by a factor of many XXX. Why do you think it sales?


Then please explain me why doesn't even the 1066MHz FSB limit GPU performance on Intel machines.

Again you are thinking of performance. I wouldn’t let you do any CPU or any Highway for example. You don't know why the connecting bus speed is always higher than the device speed.
Bus speed in any technology can never be the bottleneck, go see Ati Xenos design on the XBOX360:
ATI Xenos: Xbox 360 Graphics Demystified

HDD speed < Sata Interface speed
AGP card performance < AGP Interface speed
PCIe card performance < PCIe Interface speed
Processor performance < Bus interface speed
Cars travelling in a Highway < Highway size


Even today you must be looking really hard to notice speed differences between 8x and 16x PCIe.

If you didn’t understand the last answer, you don’t know the answer to this one either.

Scientia from AMDZone said...

ho ho

"I'm curious, what metrics do you use to say that AMD develops their process further than Intel?"

I doubt Intel will make it over 3.0Ghz on 65nm. It appears AMD will.

"Pentium M was rather nice at 90nm."

No. Pentium M drew more power than Turion even after it was upgraded from Banias to Dothan. Higher power draw than Turion but no additional performance than Turion pretty much equals worse.

"Battery life, perhaps, but most certainly not performance."

Again, this isn't a desktop system. Battery life is most important.

"... and Nehalem is faster than Penryn :)"

while Bulldozer is faster than Shanghai.

" Also I'm quite sure AMD will have to lower its K8 prices again this year. If not sooner then after Penryn comes and pushes current Core2 prices down."

Penryn won't be out in any real volume in Q4 so it won't have that much effect on prices.

"My best guess would be around 2009/2010 if everything goes fine. If not then perhaps a year later."

Maybe if you are only talking about replacing PCI-e and CSI with a fiber link. However, AMD could do the same. I would guess AMD has to be looking at this for HT 4.0 or 5.0. Afterall, this has been used for more than two decades in telephone and LAN.

"there are four 690 based motherboards listed on estonian pricewatch site. Against those four we have 25 motherboars with 965 chipset."

The 965 boards have been out a lot longer. NewEgg shows 42 965 boards. I would say 35 versus 42 is excellent.

"Of cource they have. Going from zero to something is a massive gain."

Sometimes you say things that make it sound like you don't know anything. AMD had mobile chips before Turion. AMD also had machines on the TOP 500 supercomputer list before Opteron.

"Intel has two teams working on x86 in tic-toc mode pumping out new cores and refreshes every two years. AMD Pipe seems to be something similar. Will it have additional two teams working on its mobile CPUs?"

Pretty much. The work on Turion, MIPS, and Geode will be consolidated for Bobcat.

"Though it should be known that mostly people use their laptops just the same way as desktops: connected with wires."

My desktop replacement functions essentially as a portable but not as good for mobile. You can also do desktop replacement with low power versions of the desktop chips. The desktop space will shrink anyway with DTX and mini-DTX cases.

Scientia from AMDZone said...

aguia

"I have already showed that Intel demo Conroe at 3.5Ghz and Kentsfield at 3.33Ghz and none of this CPUs still exists today.
Intel and AMD high clocked CPUs, keep dreaming!"


You need to understand the difference between overclocking which the Intel chips were and normal speed which the 3.0Ghz AMD chips were. Find something that says that Intel demonstrated those chips with stock HSFs like AMD did and you'll have a point.

Scientia from AMDZone said...

abinstein

"The main reason is MCM with HT will inevitably result in a NUMA system, which most desktop applications don't optimize for. "

This would only happen if AMD completely separates the two memory channels and I can guarantee that this is not the case. Using the same archicture AMD has plans to offer memory mirroring.

enumae said...

Abinstein
In other words, AMD's claim that Barcelona is superior than Xeon would be even more true if the test was one of the two possible SPECfp sub-tests.

You are really funny. I am not trying to say that one is better than the other, just that it looked strange.

Yes the question was clearly there but I deemed it not worthy of a reply.

Ok

Please, just read my comment and tell me how does this sub-test mislead any conclusion?

Because they come across as though it is a full SPECfp test, like I said you had thought it was.

Barcelona is having much better fp than both Clovertown and K8.

There is no question it is much better than Clovertown.

There is the question about it being much better than K8, we will have to wait and see.

enumae said...

Abinstein
You are comparing 3.0GHz K8 with 2.66GHz Core 2. At the same clock, dual-socket K8 X2 is just about 10% faster than dual-socket Core 2 Duo in SPECfp.

Well don't I feel dumb... Sorry about that.

Scientia from AMDZone said...

lex

"Give you credit you write like"

I can't give you the same credit. You write a lot of verbage but say very little. You also seem to have some trouble with English and are responding to things I didn't say.

"Some think you just go to Applied, ASML and put those tools in the fab and viola you got high performance, high yielding silicon just like that."

Who thinks that? I remember when ATI complained that although TSMC and UMC had similar processes it took a bit of work to port to UMC.

"As such AMD will struggle mightily at 45nm and 32nm."

Based on what besides your (obviously biased) opinion?

" Neither will give them equality to the INTEL but their disadvantage is actually growing."

Even though nothing indicates that. It's a secretly growing disadvantage apparently.

"They haven't put the necessary leg work in for the past few years"

So, you found out that AMD's 77 techs at East Fishkill do nothing but play poker all day? That was supposed to be a corporate secret.

"if they think they can pull it off with the spending limits they have."

Which are proportionate to Intel's.

"I'll just address a few of your comments."

As well as ones I didn't say.

"WRONG, if 90 and 65nm are any indication AMD/IBM process has been inferior to INTEL offerings."

Let's see. AMD's 90nm used less power and clocked higher. It also appears that AMD's 65nm will clock higher as well with similar power draw. I guess I'm missing the part where AMD is worse.

"it will be a penalty them by either 2x in gate leakage or 20 percent in raw performance."

Or not. AMD seems to be indicating 20% more performance at the same power draw. Now, this is worse than Intel which is reporting both more performance and reduced power draw on 45nm.

"a significant lithography change? That sounds very risky to me actually."

It is risky if you don't approach it the right way.

" INTEL seems to have gotten by with tried and true dry lithography."

Right, Intel wasn't up to it.

"I believe if AMD does ramp as they claim they will they will be the first to go to production with immersion "

No. TSMC is going to immersion late this year.

"I have no confidence that AMD has the resources, deep pockets, track record as a process leader"

What a surprise after all those earlier compliments for AMD.

"nor the most important thing TIME to do this."

Well, I guess AMD wouldn't have time if they hadn't been working on immersion for the past three years. Oh, that's right; you wouldn't believe that since you've already claimed that AMD has put no effort into process technology in the past several years.

"There is NOTHING to be gained by a small fab."

I didn't mention having a smaller FAB. AMD will have an advantage by having two similar FABs at the same location. Intel's disadvantage is having widely separated FABs.

"The technology problems are the same whether you are running a few thousand wafers a week with a couple tools of each type or much larger scale with even more tools."

Exactly, which is why AMD will be able to ramp FAB 38 more easily.

"Thus a small fab is the worst thing for AMD, they need a Mega fab to maximize resources and distribute the fix cost of the factory."

Exactly, which is why FAB 36/38 will be one big FAB. I belive it will be about 3X the size FAB 30 was in 2003 but will produce about 4.5X the wafer area per month.

"When do 32nm R&D and manufacturing costs start?"

What does R&D have to do with the temporary expenses of building FAB 36 and merging with ATI? R&D is obviously an ongoing expense.

"That is where Intel’s tick tock really will crush AMDs design/process conversion cadence."

A two year cadence will crush a two year cadence? Interesting.

"don't tell me about YOY growth when you are starting from a miniscule number"

Do you have trouble understanding scale? If AMD's volume is miniscule then I guess Intel's would be small or maybe medium. Let me help you. Compared to Intel, VIA's CPU volume is miniscule. AMD's volume is closer to Intel's than VIA's.

"The bottom line is AMD MS encroachment have been stopped and if anything its all going back to INTEL across most segments."

This must another one of those secret changes that the numbers don't show.

"Sorry unless they find a whiteknight AMD is finished,"

This is quite a surprise after your earlier positive comments about AMD.

"they don't have the cash flow to fund the technology they need to compete."

This would be the cash to fund those 77 techs who play poker all day? Or would this be the $1.8 Billion that AMD is spending on tooling for FAB 36?

"wait and see what happens end of next year when 45nm is suppose to happen for AMD . "

It's supposed to happen earlier than that.

Aguia said...

So, you found out that AMD's 77 techs at East Fishkill do nothing but play poker all day? That was supposed to be a corporate secret.

LOL. If that’s the case I'm qualified to join the team. ;)

abinstein said...

enumae -
"Because they come across as though it is a full SPECfp test, like I said you had thought it was."

The problem is really that we didn't attend the event in person, and we had to watch it through a video.

Had you or I were there we could've asked the lady which SPECfp benchmark it is running, then we'd probably have got a good answer, or at least is acknowledged the benchmark is not full SPECfp.

abinstein said...

Ho Ho -
"Then please explain me why doesn't even the 1066MHz FSB limit GPU performance on Intel machines."

Apparently GPU performance does not take much FSB bandwidth. Is it that difficult to understand? Do you load/store every pixel with x86 instructions when writing OpenGL or DirectX?


aguia -
"You don't know why the connecting bus speed is always higher than the device speed.
Bus speed in any technology can never be the bottleneck"


Ho Ho, aguia's reply was absolutely correct, and you should learn some queueing theory and computer microarchitecture before making other erroneous comment like you did.

In any queueing system (memory be an example), utilization, or arrival rate over processing rate, should always be less than 1 or you get infinite queue length (stalled processor). The performance of the system is measured not by bandwidth, but queueing delay.

A higher FSB utilization will drastically increase the access delay. Intel's adding more and more cache is not to increase memory bandwidth, but to reduce memory latency; so is its raising FSB speed.

enumae said...

Abinstein
Had you or I were there we could've asked the lady which SPECfp benchmark it is running, then we'd probably have got a good answer, or at least is acknowledged the benchmark is not full SPECfp.

I can accept that.

Thanks.

InTheKnow said...

Scientia said...

AMD will have an advantage by having two similar FABs at the same location. Intel's disadvantage is having widely separated FABs.

You are completely mistaken here.

First Intel has a virtual factory system that ties all of their factories on a given node together through copy exactly. With the aid of email, an intranet system and a little invention called the telephone there is very little given up due to lack of proximity. And Intel's virtual factory system is larger than your proposed AMD mega fab at each 12 inch technology node, 90nm, 65nm and soon to add 45nm.

Second, Intel has multiple fabs on at least two of their campuses. In Oregon they have D1D, D1C and Fab20. In Arizona they have Fab12 and the ramping F32. New Mexico is supposed to be shutting down F11, just leaving F11x, so they don't count going forward.

I can already here the cry "but they are on different technology nodes"

To which my reply is "Yeah, they are. So what?"

A good 95% of the tooling is identical. And if you know anything about HVM process technology, you know that almost all of the issues boil down to equipment issues. Once you are in HVM, process parameters are pretty well fixed. You aren't looking for the right combination of pressure, temperature, time, stoichiometry, etc. You are just trying to figure out how to keep your tool up and producing high yields. And for that, a guy with the same toolset is still another resource, no matter what technology node he is on.

An engineer doesn't forget all the learnings from the last technology node just because the process changes. The toolset still behaves the same way it did on the last node and presents the same (generally mechanical) challenges.

Trust me on this. The "advantage" you perceive AMD gaining from the physical proximity of the two fabs is non-existent. I wouldn't presume to lecture you on the finer points of computer science, please show me the same courtesy when it comes to HVM process engineering.

Unknown said...



When R600 was reviewed it was on an 80nm process with leaky 80hs transistors.


Trying to blame TSMC's process here? G80 is the fastest, most advanced, GPU in the world and it's produced at 90nm. AMD's GPUs are broken; not TSMC's process.

Aguia said...

Trust me on this. The "advantage" you perceive AMD gaining from the physical proximity of the two fabs is non-existent.

Really?!
How having the same employees on both factories and going from factory to factory as the needs arise practically "on foot" is not an (huge) advantage?

I also don’t think that some problems or most of them as you implied, can be solved on the phone, or on the mail, PDF, ... If that’s the case then AMD as an huge advantage over Intel, their power point presentation is great!

InTheKnow said...

aguia, have you fixed many tool issues? If so then you know that 75% of the process doesn't have to happen in the fab. You do data analysis on the failure mode, check tool parameters and develop a troubleshooting plan. Then you collect the data from the plan and determine the next steps until you have root cause. This part of the process can be done from anywhere.

Now the other 25% of the process is based on observation and physical proximity does help for that. And since Intel has multiple fabs on most of their larger sites they can have someone "walk down the hall if they need to." Or did you not pick up that little detail from my earlier post?

Aguia said...

So if that’s the case InTheKnow then that could explain the why AMD achieved (slide 22) 65nm at mature yields so fast, or this is unrelated?

2007 Analyst Day

abinstein said...

"75% of the process doesn't have to happen in the fab."

Sure, you don't need to do all the work on-site. However, those work are still done more efficiently with physical proximity to the fab.

abinstein said...

intheknow -
"First Intel has a virtual factory system that ties all of their factories on a given node together through copy exactly...

Second, Intel has multiple fabs on at least two of their campuses..."


Your statements actually show Scientia's right, that having fabs spread around is a disadvantage. That is precisely why Intel needs to copy-exact, to create "virtual fab," and to build multiple fabs on each campus.

"An engineer doesn't forget all the learnings from the last technology node just because the process changes."

What the engineer doesn't forget is probably not much useful to the next node. It is a know fact that from 90nm on "disruptive" techniques are used across process nodes. I think even Intel acknowledged it, that they have to apply very different tricks to each new process node, that there is now a learning curve on every new process.

"Trust me on this. The "advantage" you perceive AMD gaining from the physical proximity of the two fabs is non-existent."

I believe not even Intel engineers would trust you on this, or they wouldn't want copy-exact, nor building multiple fabs per campus. They would just have built fabs on the cheapest lands and use your wonderful inventions such as e-mail and telephone to communicate with each other.

Scientia from AMDZone said...

intheknow

Intel realized this a long time ago. Intel's first development FAB was D1 which later become FAB 15.

The second was D1B. Intel realized the advantage of having everything together so it built RP1 and D1C next to D1B. RP1 is where process research takes place. D1C replaced D1B as the development FAB and D1B became FAB 20. Intel later replaced D1C with D1D and D1C became a 300mm production FAB. So, in one location Intel has everything from process research to production research and development to full production.

I assume that Intel maintains FAB 20 so that it has a 200mm facility on hand as a pathfinder for the other 200mm FABs. Presumably once all of the chipset production is moved to 300mm FABs and other technologies are either divested or moved to 300mm Intel will probably upgrade or annex FAB 20.

Intel has also adopted this pattern of larger facilities by creating double FABs at Chandler AZ, Leixlip Ireland, and Rio Rancho. You are mistaken about F11. Intel rennovated as much of the old FAB 11 as it could and this was added to FAB 11X. There is a similar situation in Israel with FAB 18 but I believe Intel decided not to rennovate it.

AMD gains an advantage because nearly 100% of its production is in the same place. It's true that you can email, talk on the phone, and even teleconference but I also know that organizational overhead can be significant. Having everything in one location removes some of that overhead and makes things more efficient. You also cannot share personnel between Rio Rancho and Leiplix very easily.

AMD needs two of the same type of FAB though. Right now the wafers and tooling between FAB 30 and FAB 36 are different. AMD needs FAB 30 to be the same 300mm production so that it matches the research at East Fishkill and the production at FAB 36 and Chartered.

Now, I'll admit that if AMD builds the NY FAB they'll basically be the same position as Intel. It is interesting though that AMD wants a NY FAB because of its proximity to IBM. One can speculate that this would make it easier for AMD's East Fishkill personnel to visit the production FAB.

Since this is seen as desirable by AMD we then have to assume that Intel currently holds this advantage over AMD since all of its research and devlopment facilities are in one location at Hillsboro.

Scientia from AMDZone said...

giant

I'm not sure what point you are trying to make. ATI stated that they knew the problem existed with the 80hs transistors but didn't believe that they had time to move to 65nm. Now they are. Given the problems at 80nm, this seems reasonable to me.

Blame is not a factor. This simply concerns a problem and its solution. The R600 reviews showed the problem and it will be fixed in Q4.

InTheKnow said...

aguia said...

So if that’s the case InTheKnow then that could explain the why AMD achieved (slide 22) 65nm at mature yields so fast, or this is unrelated?

If anything, your comment supports my point. You are referring to the 65nm ramp. AMD has only ramped the one fab. It is all new tooling (12 inch) so there is no help to be had from next door.

And that plot looks a little odd. Every other plot I've seen for ramps puts time on the x-axis. This one uses wafers run. So while AMD has certainly not wasted a lot of Si during the ramp, I don't know if it was fast or slow just looking at the plot. They could have spent months processing one wafer at a time through the fab getting everything perfect on each wafer before moving on. I don't think they did that for a minute, but I can't tell any different from the plot.

InTheKnow said...

abinstein said...

Sure, you don't need to do all the work on-site. However, those work are still done more efficiently with physical proximity to the fab.

Have you ever been in a fab? It is much easier to do analysis outside of the fab. I remember having a conversation several years ago with and engineer at Fairchild who was saying how much nicer it was to work in the new fab (it had data acquisition systems that the older fabs didn't) precisely because you didn't have to run into the fab every time you needed data from the tools.

And please note that I did say it was advantageous to to have access to the fab about a quarter of the time. But you hardly need 18 people packed around a tool to watch it. 2 or 3 is more than enough. So you can have your hands on guys local and they can interface with anyone they need to outside of the fab in any location.

Your statements actually show Scientia's right, that having fabs spread around is a disadvantage. That is precisely why Intel needs to copy-exact, to create "virtual fab," and to build multiple fabs on each campus.

Oh there is an advantage to a campus with multiple fabs alright, but it has nothing to do with solving process issues.

Intel is a big company with plenty of bright people working there. If the disadvantages of the virtual factory were that large, don't you think they would have concentrated all of their like fabs together. Instead, they chose to ramp D1D on 65nm, followed by Arizona, the Ireland, then D1C. Arizona was a new site. They have plenty of room for expansion at the Oregon site, and yet they chose to build elswhere (F12 in Arizona). What would you suggest the reason for that is?

What the engineer doesn't forget is probably not much useful to the next node. It is a know fact that from 90nm on "disruptive" techniques are used across process nodes.

They may use "disruptive" techniques, but the entire process isn't disrupted. Let's take Intel's 45 nm process as an example. They replaced Silicon oxide gates with metal gates. This changed a handful of process step. So you are right, there is no past experience there. However for the majority of the line, there are no significant process changes. All that learning from 90nm and 65nm still applies.

I believe not even Intel engineers would trust you on this, or they wouldn't want copy-exact, nor building multiple fabs per campus. They would just have built fabs on the cheapest lands and use your wonderful inventions such as e-mail and telephone to communicate with each other.

For the record, the intel engineers I know don't "want" copy exactly. They just have to live with it. And they build multiple fabs per campus for cost reasons, and not just land. Shared facility costs represent a huge savings.

You have failed to present any evidence to refute my claims just made random objections that I have nonetheless tried to demonstrate the flaws in. Feel free to bring some evidence to the table next time.

InTheKnow said...

Scientia, let me be perfectly clear here. Without a doubt there are advantages to having multiple fabs on a single campus. I've never disputed that. However, those reasons aren't for process improvement or sustaining. It is all about dollars. The infrastructure needed to sustain a fab is extensive and expensive. When you can put multiple fabs on a campus, you save big bucks by spreading out those costs between all the fabs. The cost to add capacity to the gas pad for example doesn't double when you add in a second fab. The cost increases but is far less than double. That observation applies to all the peripherals that surround and support the fab.

Intel has room to build out the campus around D1D. If the advantages of spreading out their facilities didn't outweigh the disadvantages, I would expect them to build out all their campuses to their maximum occupancy before building up a new campus. Intel has the options and the experience to choose between them. I have to assume they know what they are doing or they would have gone the way of the dinosaurs by now.

abinstein said...

intheknow -
"And please note that I did say it was advantageous to to have access to the fab about a quarter of the time."

This advantage of "physical proximity" is stated in your original statement, and is what I was referring to.

"Oh there is an advantage to a campus with multiple fabs alright, but it has nothing to do with solving process issues."

Maybe I misunderstood your discussion. I didn't try to claim multiple fabs per campus can solve process issues faster. I was saying precisely that it costs less to not have fabs spread around the globe.

"For the record, the intel engineers I know don't "want" copy exactly. They just have to live with it."

We all know why copy exactly is important - so engineers can bring the experience several physically separated fabs together.

Roborat, Ph.D said...

abi said: We all know why copy exactly is important - so engineers can bring the experience several physically separated fabs together.

it is the other way around. it is to make several physically separated experience the same thing. but i'm just splitting hairs.

also several years ago intel learning that copy exactly allows them to do simulatneous multi-fab ramps with vertical yields. so CE! has gone beyond just a change management philosophy and into an NPI strategy.

Scientia from AMDZone said...

Intel's Copy Exact process works for them. It requires two or three months of calibration runs before you can run a batch for production. Then, at that point you have a dispersed production system.

However, I think someone would have to agree that what works for Intel (under one management and one organization) would not work for AMD. I just cannot imagine the combination of IBM/AMD/Chartered without APM.

Once AMD has two 300mm FABs there is a theoretical capability to run a production batch cold on the second FAB without a calibration run. However, I'm not sure that this would ever happen. It seems like that while you were running test chips on the first FAB you could calibrate on the second as well. However, test masks are only changed quarterly so I suppose it could potentially happen.

APM also has a theoretical ability to adjust batches in flight which I don't believe is something that CE can do. Again, I'm not sure how often this would matter in practice. In theory, it could help you save a batch that was not quite what you wanted or potentially adjust a batch for higher speed or greater power savings. However, there are limits to how far you can push a batch once it is already in process.

Scientia from AMDZone said...

I wonder if this is only temporary. Kubicki claims that:

all the good Intel deals went out of stock in less than 24 hours.

Vendors who restocked quantities (Newegg included) returned pricing to pre-cut levels almost immediately.


So, I'm wondering if these good prices were just for discontinued stock or whether they will come back down again.

enumae said...

Scientia
So, I'm wondering if these good prices were just for discontinued stock or whether they will come back down again.

Well all you have to do is look at the current pricing from the respective sites that are in the Dailytech Blog article...

1. MWave: $282
2. ClubIT: $295
3. ZipZoomFly: $295
4. NewEgg: $309
5. E-Wiz: $279

Prices shown are as of 8/3/2007, 7:30PM

Aguia said...

More info about Barcelona and future AMD products

AMD Tech Day at Develop 2007

greenmachine said...

Scientia,

I think everyone may find this thesis of interest.

http://etd.lsu.edu/docs/available/
etd-06122007-093459/unrestricted/
Prakash_thesis.pdf

I am sorry to break up the html link, but I could not get it to fit on one line.

Perhaps you could write an article on it. It seems to verify a lot of your previous articles on the strengths and weaknesses of core 2 duo and amd architectures.

I too have been looking for valid benchmarks instead of just fog. I was surprised at how close the two are in most applications.

Thanks

PS I originally posted this in the Q2 Earnings article erroneously. Forgive the repost

Scientia from AMDZone said...

Fuad is showing 45nm penryn prices. So it looks like these will be what AMD has to deal with:

X5460 3.16 GHz, 12 MB, 1333 FSB, $1,164

E5450 3.0 GHz, 12 MB, 1333 FSB, $852

E5440 2.83 GHz, 12 MB, 1333 FSB, $695

E5430 2.66 GHz, 12 MB, 1333 FSB, $467

E5420 2.5 GHz, 12 MB, 1333 FSB, $332

E5410 2.33 GHz, 12 MB, 1333 FSB, $273

E5405 NA, 12 MB cache, NA, $229

The best guess for E5405 is 2.16 Ghz.

AMD is likely to have 2.5Ghz in Q4 so I would guess they would match the E5440 2.83 GHz. So, that would be something like:

K10 quad 2.5Ghz - $695
K10 quad 2.3Ghz - $467
K10 quad 2.2Ghz - $332
K10 quad 2.1Ghz - $273
K10 quad 1.9Ghz - $229

AMD would need 2.7 and 2.8Ghz to match the top two grades.