Friday, January 26, 2007

AMD Q4 Earnings -- Good Or Bad?

AMD has delivered its Q4 2007 revenue numbers. The results can be pretty much as expected, disappointing, or evidence of a company going down in flames depending on how distorted the analyst's view is. Lately, the latter opinion seems to be the most prevalent with many articles mentioning the $574 Million loss. However, this view is, for the most part, inaccurate.

I think everyone with common sense expected that when AMD bought ATI that it would lose its Intel contracts. This is certainly what I expected. This loss of contracts created a big loss of revenue for ATI in the fourth quarter which again was expected. That it would take some money for this merger to be completed was again very much expected. I've seen a few people wail about the loss of the Intel contracts and make hasty projections that the ATI section will never again turn a profit and will only drag AMD down. Usually, this idea is lumped with the idea of AMD's several Billion debt and the projections then begin running to sale of ATI or even bankruptcy. These projections are something of a source of amusement but not much more. Let's see if we can look at this more realistically.

Intel has a big weakness in the area of motherboard integrated graphics where they are today almost unchallenged. ATI brings this capability to AMD in a way that is profoundly beneficial for both companies. It is this failure to recognize the benefits that is the most surprising to me. AMD benefits ATI tremendously because it has more money for R&D than ATI has. ATI also gets the benefit of early notification of AMD's technological directions. These two things are bound to give ATI the boost it needs to be more competitive with nVidia. ATI likewise gives AMD an all in one solution for OEM's who prefer this to putting together systems piecemeal. AMD will be able to deliver packages where the southbridge and graphics chips bear the AMD logo rather than just the cpu. This should give AMD a tremendous boost in corporate sales. This boost in turn provides a ready market for ATI's products which should easily make up for the lost Intel contracts.

Some people have suggested that AMD should have used its cash to build a third FAB rather than purchasing ATI. That does not seem reasonable to me. As a business gets larger some customers will expect more services to be provided and it is very likely that AMD had went as far as possible while providing only processors. This not only means that AMD can provide complete solutions but it also means that AMD can guarantee support when it is needed instead of when 3rd parties decide to provide it. This is something that would have done AMD some good as far back as the introduction of the K7 and would have done even more good when K8 was first introduced. Rather than trying to twist this around to some kind of problem for AMD, people should see it for what it is: the loss of yet another of Intel's previous advantages. AMD broke out of its single FAB limitations in 2006 and this lays the groundwork for broad platform support in 2007. I doubt there is anyone at Intel who while proclaiming the introduction of Core 2 Duo doesn't recognize the great strategic losses to AMD during 2006. Overall, AMD gained more with FAB 36 and ATI and than Intel gained with C2D.

For those who have missed this, it only takes a moment to look back at history. The progress that AMD made with K7 up to 2001 was largely pushed back in 2002. And, K7 was unable to make any real headway all through early 2003. With K8, however, AMD's volume share has grown 50% and Intel has shown no signs of being able to take this back. With a second FAB and ATI it is even less likely that Intel can take back any volume share in 2007. Let's look at the numbers:

For processor revenue from 2005 to 2006:
Intel shrunk 15.3%
AMD grew 34.6%

For processor revenue from 2004 to 2006:
Intel shrunk 2.5%
AMD grew 201.9%

I think anyone with common sense can see that Intel has slipped approximately 2 years in terms of processor revenue parity with AMD. I'm sure that Intel will have recovered to 2005 levels by the end of 2007 however I'm also certain that AMD will grow at least as much as Intel during the same time period. Of course, Intel enthusiasts will want to concentrate on just the Q4 numbers which seem to be the only bright spot in a very bad year for Intel.

If we ignore the acquisition costs then AMD's earnings were less than half of Q3's. Since AMD increased volume by 9% from Q3 this has to indicate a reduction in Average Selling Price (ASP). This would be the effect from Core 2 Duo. AMD's inventories are up 1% from Q3 which appears to be due to ramping of FAB 36. Intel's inventories are down very slightly. Q1 should provide a better inventory picture. Intel is slightly better off than AMD in terms of cash although AMD has about twice as much debt as cash while Intel has about four times more cash than debt so overall Intel is much better off. Compared to Q4 2005 Intel is down 11% while AMD is up 3%. Intel does appear to have gained back 0.8% of revenue share. However, with the boost in volume in Q4 this is probably only a temporary change.

Basically, Intel's numbers show that they are back on track and slowly growing again from the large losses earlier in the year. AMD's numbers show that they are trying to hold onto volume share and are currently willing to give up revenue to do that. This has been a drop for AMD of about 10% on ASP. This 10% accounts for the lower earnings compared to Q3 06. The outlook is also pretty boring. Obviously, Intel will continue to ramp C2D and not much more than that until late 2007. Then they will release a quad FSB northbridge allowing Woodcrest to move into the 4-way area and they begin slowly releasing 45nm. AMD will continue to ramp FAB 36 although this is partially offset by beginning to ramp down FAB 30 for the 300mm conversion. They will introduce K8L Barcelona in Q2 which should make 2007 K8L's year in the same way that 2006 was for C2D. AMD's margins should be up in Q1 and their channel distribution should be fine. The clearest signal from 2006 however is that the primacy of the desktop is fading fast. Prices are down for desktop compared to mobile and server and clearly these will be the areas that AMD will pursue. Likewise, AMD is growing faster in terms of commercial than consumer. Most of this stuff is pretty boring.

I see three battles taking place in 2007 that are not typically talked about. The first is with instruction extensions to x86. Up until AMD64, Intel was able to create new extensions which immediately became the standard. The best example of this is when AMD's 3D Now was brushed aside by SSE. However, it should be interesting to see whether or not Intel adopts AMD's planned instruction extensions in 2007. This is particularly noteworthy because adoption of these by Intel would tend to hurt Itanium. The second question is memory. It is no secret that Intel pushed JEDEC towards DDR2 at the expense of DDR. And it is no secret that this worked to Intel's advantage because the higher latency was offset by Intel's larger cache. However, it appears that AMD may be able to get JEDEC to extend DDR2 to faster speeds rather than just moving to DDR3. This would tend to be disadvantageous for Intel. Finally, it remains to be seen what the review sites will do during 2007. It is almost a certainty that by 2009 the typical pro-Intel stance will be broken however it is not clear when this might happen. It is somewhat possible that reviews could become more professional (such as not using the Intel Compiler) by the end of 2007 when head to head testing for quad cores should reach a peak. However, it is also possible that sites like Toms Hardware Guide and Anandtech could feel that Intel will retain its dominant position or make a strong comeback with 45nm. If this is the case they will continue to hold out in Intel's favor until sometime in 2009.

127 comments:

enumae said...

First, I just would like to say that was a really good article...

There may be many variables that have not been factored in, and the biggest one being Intels transition to 45nm and the abillity to keep lowering prices while adding more cores, but I'll touch on that at the end.

Intel has a big weakness in the area of motherboard integrated graphics where they are today almost unchallenged.

Pleas explain how and where they have a weakness. Almost everything I have read is stating that Intel is capable of Vista Ultimate and HD playback on the IGP's. They have no need to become a graphics company (though some reports say they are trying), they are making what the market needs, granted it may be just enough, but the prices are cheap and the products capable.

This should give AMD a tremendous boost in corporate sales. This boost in turn provides a ready market for ATI's products which should easily make up for the lost Intel contracts.

They should, but just to touch on 45nm and Intels ability to provide a platform. 45nm will be cheaper than 65nm, Intels chipsets in general move to the next processing node, and having the FAB's has to help price reduction more than outsourcing to Chartered or TSMC.

As a business gets larger some customers will expect more services to be provided and it is very likely that AMD had went as far as possible while providing only processors.

I think that is wrong, AMD could have easily sat back and enjoyed the profits, saved some more money/got out of debt and then aquired ATI, they got too agressive and it may end up hurting them in the short term.

The aquisition of ATI really had to upset Intel, but Intel knew everything that we are seeing now about the aquisition charges and fees, and knew they would pull ATI's business as did AMD.

What AMD did not see is Intels ability to sell Cloverton at around the same price as Woodcrest, and also what kind of performance it would have compared to a 2P server/workstation Opteron system.

...the loss of yet another of Intel's previous advantages.

But when will this show up and at what cost to AMD's finances for the near term?

Let's look at the numbers...

There looks to be two mistakes, Intel shrunk 12.4% from 2004 to 2006. AMD would have grown 134% from 2005 to 2006 so that it matches your 201.9%.

I'm also certain that AMD will grow at least as much as Intel during the same time period.

From the numbers I had looked at ATI's margins were about 35%, while AMD was about 50-55%, if you were to factor those into the equation AMD's margin should be around 42.5-45%, that's quite alot of money.

They will introduce K8L Barcelona in Q2 which should make 2007 K8L's year in the same way that 2006 was for C2D.

Maybe it would if they had a year between launches, but being 45nm should be coming out about Q4 people may wait. If K8L is as good as they claim and no benchmarks from Intel are floating around, then your theory is sound, but thats still months away.

Also looking at the news today about Core 2 Duo's, the 45nm version will be getting Hyperthreading. I am not sure if thats good or bad, what do you think?

However, it should be interesting to see whether or not Intel adopts AMD's planned instruction extensions in 2007.

I am having a hard time finding them, could you link to the instructions? Thanks.

This would tend to be disadvantageous for Intel.

Why? Doesn't DDR3 have lower power requirements than DDR2 at 1066 and above and wouldn't this also help AMD and the end users?

It is almost a certainty that by 2009 the typical pro-Intel stance will be broken however it is not clear when this might happen.

This doesn't make any sense, almost certain, not clear?

Just giving you a hard time, I reallly enjoyed the article.

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

Now my take of 2007 and entering 2008 is that it will be all about quad cores in Servers.

If Intel is able to have 45nm quad cores with lower or equal power to 65nm dual cores, this is where I see Intel hurting AMD's prices, and where AMD's limited capcaity hurting them.

It is no secret that Intel has alot of manufacturing muscle and being at 45nm first will let Intel dictate prices and in general AMD's margins.

Again, great article.

Red said...

It seems to me that most analysts are slamming AMD on their "positive" earnings, especially on the gross margin. The -$0.5B just makes for nice headlines.

AMD will be able to deliver packages where the southbridge and graphics chips bear the AMD logo rather than just the cpu. This should give AMD a tremendous boost in corporate sales.
Yes, let's ring in the cash registers! Those AMD logo will sell like hotcakes! Intel's IGP does good enough for the market segment that it is in (bottom of the barrel PC's). And Intel has motherboards.

Overall, AMD gained more with FAB 36 and ATI and than Intel gained with C2D.
Good for Fab 36, else they be sub 40 gross margins by now. I have high hopes for R600, but so far, I don't really see what they've gained, unless you think that gamable IGP's on $250 PC's is important.

I'm also certain that AMD will grow at least as much as Intel during the same time period.
Well, that's kind of a given. And if they'll grow at least as much as Intel, that's kind of bad. So you don't share Ruiz's view of 20% growth?

http://seekingalpha.com/article/24929
According to AMD, ASP from desktops are "slightly down".
Server processor unit shipments were essentially flat quarter over quarter, and ASPs declined significantly.
First, I would like to remind you that our ASPs in the server space were higher than Intel’s, so there was a correction, certainly, as they came with a more competitive product.

They will introduce K8L Barcelona in Q2 which should make 2007 K8L's year in the same way that 2006 was for C2D.
Not really. AMD can't ramp as fast/much as Intel for the effects to be felt. And we have no idea how it'll perform. 40% better overall, 40% better than Clovertown, 60% performance per watt, etc.

It is almost a certainty that by 2009 the typical pro-Intel stance will be broken however it is not clear when this might happen.
Yes, that "typical pro-Intel stance" stench was horrible during the X2 days, wasn't it?

Happy Almost February!

Greg said...

Enumae and red, what you fail to factor in is that Intel JUST made their IGP chipsets that are finally compatible with Vista, while ATI's has been out for near eternity and been compatible.

While the end user wont appreciate this, Dell, HP, and the like will love it because their customers wont complain when the newest OS comes out and they can't use it, or other things like that. While it may seem like that's not HP and Dell's problem, they are paying people to answer the phone, to answer stupid questions, that the customer should have known about. To summarize, by making their integrated graphics stronger, PC companies further idiot-proof pcs, thus increasing margins, by decreasing support costs. Also, Intel obviously does feel threatened by the lack of its own GPU business, as it seems to be pushing its own GPU group (larrabee) pretty hard.

If anything, adding hyperthreading seems like a bad idea, as with the dual-core chips, it simply caused the fsb to be saturated far too much.

While holding off on ddr3, despite its lower power costs and higher bandwidth may seem like it's bad for the customers, it's really not if they're AMD's customers. The question of importance here is how much performance is power savings worth, especially if, as far as we can tell, increased bandwidth to the memory yields very little in terms of performance.

I think it very likely that 2007 will be k8l's year, as 45 nm will only allow Intel to increase core speeds, not necessarily bandwidth, decreased latency, or more IPC. They can get more cache, but the performance improvements from cache will eventually hit a wall, and to truly take advantage of them, there will need to be architectural changes. While they will have a higher clockspeed than AMD processors, AMD seems confident that they have higher performance per clock, and AMD has shown that they obviously are able to milk a process for all its worth with their use of newer transistor tech. Also, Intel's cache advantage will go away when AMD introduces L3, and faster cache interconnects.

Enumae, while Intel's production advantages are large, they're production advantages that Intel has always had, and that doesn't seem to have made that huge of a difference ever. AMD production disadvantages, however, are disappearing, and that would seem to me to be their biggest problem.

enumae said...

Greg
Enumae and red, what you fail to factor in is that Intel JUST made their IGP chipsets that are finally compatible with Vista, while ATI's has been out for near eternity and been compatible.

Thats strange, because the Intel 945G chipset has been around for quite some time and it is aparently Vista Premium capable. Here is a link to a computer being reviewed with Intels 945G chipset in June of 2005. This link is just to show how old the Vista capable chipset is.

Here is a link to some videos of Vista running on a 945G chipset, or just search YouTube so this is hardly JUST.

Enumae, while Intel's production advantages are large, they're production advantages that Intel has always had, and that doesn't seem to have made that huge of a difference ever. AMD production disadvantages, however, are disappearing, and that would seem to me to be their biggest problem.

AMD is currently only at 30% dual cores and they were capacity constrained, now in 2008, late 2007 the mix for servers will be quad core, so how ever much you feel there disadvantages are disappearing, they will surely resurface in the early part of 2008.

Remember FAB 36 and FAB 38 will not be at total capcaity until the end of 2008.

Erlindo said...

Greg wrote: While holding off on ddr3, despite its lower power costs and higher bandwidth may seem like it's bad for the customers, it's really not if they're AMD's customers. The question of importance here is how much performance is power savings worth, especially if, as far as we can tell, increased bandwidth to the memory yields very little in terms of performance.

Not only that, but also the added latency. DDR3 will be running CAS latencies of 8 (that's insane)


enumae wrote:
Thats strange, because the Intel 945G chipset has been around for quite some time and it is aparently Vista Premium capable. Here is a link to a computer being reviewed with Intels 945G chipset in June of 2005. This link is just to show how old the Vista capable chipset is.


One thing is vista capable and the other is DX10 capable, which the 945G lacks. ;)

Joerg said...

I tend to tell my collegues in discussion: "Hey, they play the leap frog game in an extreme style. Welcome them to the game" And with such an volatile customer base such earning annoucment are no wonder. When you leapfrogged you compettion, you can dictate the prices ... good quarter ... when you were leapfrogged ... bad quarter.

Ho Ho said...

erlindo
"One thing is vista capable and the other is DX10 capable, which the 945G lacks. ;)"

Oh please tell me where I can buy a motherboard with DX10 capable IGP.

Fujiyama said...

Despite of positive and negative results from Q4 AMD tends to lower profitability. Intel pays for fabs, more people and more marketing of course - but earns money. AMD profits are much smaller.
The good thing is that Hector, Dirk and other AMDiers are aware of the primiary reason for that - the revenue. To small to match Intel game.

So Q4 units shipment increase, ATI merger and overall strategy should help in reaching 30% in 2008.

Scientia from AMDZone said...

enumae

Intel's weakness is that they have never had competition. It will be very difficult for them to prevent AMD's taking a share of this market. I guarantee AMD's offerings will be competitive.

And, I'm sorry but you are incorrect about costs. Intel needs both a north and southbridge so even with a smaller process their costs will not be lower.

Also, Intel cannot genuinely seel Clovertown for the same price. This would be a loss leader.

There looks to be two mistakes, Intel shrunk 12.4% from 2004 to 2006. AMD would have grown 134% from 2005 to 2006 so that it matches your 201.9%.

No, I'm only looking at cpu revenues. The numbers are correct. I might start comparing chipset numbers after Q2. Wikipedia lists the new instructions:

K8L

Look for LZCNT.

Hyperthreading for Intel C2D sounds suspiciously like a latency problem.

AMD could not have waited to purchase ATI. Without ATI they would be far less competitive in 2008. Investment.

Scientia from AMDZone said...

enumae
Remember FAB 36 and FAB 38 will not be at total capcaity until the end of 2008.


FAB 36 will be at full capacity by mid 2008. FAB 38 won't be at full capacity until end of 2009.

BTW, Real, if you want to comment on the earnings this is the place to do it. I've mentioned AMD's poor cash position and reduced earnings because of lower ASP's but you might see something I missed.

enumae said...

Scientia
FAB 36 will be at full capacity by mid 2008. FAB 38 won't be at full capacity until end of 2009.

Maybe I should have worded it differently, but what I said is still correct in the overall sense.

I should have said AMD won't be at total capcaity until the end of 2008.

Wikipedia lists the new instructions:

Thanks. Now after looking up SSE4 it would seem that K8L's new instructions are just a drop in the bucket.

Here is a link.

Intel needs both a north and southbridge so even with a smaller process their costs will not be lower.

OK, I see your point. Does making your own chipsets at your own FAB save you money?

Intel cannot genuinely seel Clovertown for the same price.

Well I don't know what to tell you, but they have been doing it, and are aparently making money doing so.

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

Erlindo
One thing is vista capable and the other is DX10 capable, which the 945G lacks. ;)

Like HoHo said, name a chipset that has been out since June 2005 with DX10 capabilities.

Besides you don't need a DX10 IGP for Vista, just DX9L, which the 945G is capable of.

Erlindo said...

enumae wrote: One thing is vista capable and the other is DX10 capable, which the 945G lacks. ;)

Like HoHo said, name a chipset that has been out since June 2005 with DX10 capabilities.

Besides you don't need a DX10 IGP for Vista, just DX9L, which the 945G is capable of.


Please take a look at this thread from AMDZone:

AMD's RD790

Intel's IGP is no match for gaming and I really doubt that it will do fine in Vista (I'll take intel's word with a grain of salt on this one).

Ho Ho said...

erlindo, you first were talking about how bad Vista support Intel has and that it has no DX10 capable IGP's coming in the near future. Now you turned it all about gaming and said that some time in the unknown future, AMD will have a DX10 capable IGP. Please decide what are you trying to say and say it, don't spin around.

Btw, I don't know anyone who has an IGP and plays games. For $70 or less I can get vastly superior dedicated GPU. For other people who don't play games, that 945 series IGP will be good enough.

Btw, MS also has said that 965G will support Aeroglass with all the features.

Erlindo said...

ho ho wrote:
erlindo, you first were talking about how bad Vista support Intel has and that it has no DX10 capable IGP's coming in the near future. Now you turned it all about gaming and said that some time in the unknown future, AMD will have a DX10 capable IGP. Please decide what are you trying to say and say it, don't spin around.

Btw, I don't know anyone who has an IGP and plays games. For $70 or less I can get vastly superior dedicated GPU. For other people who don't play games, that 945 series IGP will be good enough.

Btw, MS also has said that 965G will support Aeroglass with all the features.


1) Nvidia and ATi will launch IGPs with DX10 support. What about Intel?? (not in the near future).

2) Intel's IGPs sucks in gaming. even an nvidia 6150 IGP and an ATi X300 can both beat intel on this. I'm saying this becasue the average Joe that buys his PC at HP or Dell (or whichever vendor) will be gaming on their PCs which unfortunately are integrated mobos.
Intel does still hold a lead in integrated chipsets, but this will change in favor to AMD and Nvidia.

3) I still have my doubts that Intel's IGP runs "smoothly" on Vista. :D

TheKhalif said...

As usual you tend to look with less "rose-colored" glasses. I feel the same thing. We differ on the Fab situation though because Fab36 can make ~3x more chips than Fab30 so if the 10,000WSPM they claim to be happening right now Fab 30 can REALISTICALLY take 1/3 of Fab30 down for 300mm. If that happens by Q2, then only 1/3 of the rest of Fab 30 needs to be online for the same amount of chips.

I'm confident that that 1/3 will be used to install the first 45nm equipment (other than Fishkill) with 300mm wafers.

This will allow them to ramp around 1/3 of the total chips to 45nm in 2008.

By then I expect to see a 65nm "stream" processor for Torrenza(UMC is ready for 65nm as is TSMC).

People just as you say are "pro-Intel." I could explain it but I'd rather not.

People are mistaking balls for lack of planning. It took me awhile to accept the acquisition but when I realized that this would allow the reference platforms they recently released I felt that it was the right move.

Vista means that R600 will be a mover while RD690 and RD790 will provide, I believe, the basis for the Barcelona "glueless" chipset.

That is the next news I expect to hear from AMD, that they have updated 8000/8131 for 8Way+ SMP.

This is very important for server customers since, as you mentioned, they love getting a "whole" package even more than corporate desktops.

Red said...

http://www.dailytech.com/article.aspx?newsid=4588
Have you not heard of Bearlake?

Being vastly inferior hasn't hurt them in the past. Again, I fail to see the significance of graphics for bottom-of-the-barrel PC's.

http://www.reghardware.co.uk/2007/01/09/amd_betterbydesign_scheme/
And it doesn't look like AMD is pushing their own platform very hard, I guess because they have no wireless/motherboards. But yay stickers.

Ho Ho said...

erlindo
"Nvidia and ATi will launch IGPs with DX10 support"

When?

erlingo
"I'm saying this becasue the average Joe that buys his PC at HP or Dell (or whichever vendor) will be gaming on their PCs which unfortunately are integrated mobos"

I don't know if average Joes are stupider in US but at least in Estonia shopkeepers tell their customers that when you want to play games, any IGP is slower than cheap dedicated GPU. I can't remember when I saw last PC withouht dedicated CPU. Btw, even at the company where we work we have dedicated GPUs even though we could happily live without them.

erlindo
"Intel does still hold a lead in integrated chipsets, but this will change in favor to AMD and Nvidia."

Want to bet that won't happen before Intel stops creating IGP's?

erlindo
"I still have my doubts that Intel's IGP runs "smoothly" on Vista"

Because ...?
Even though Aeroglass is bloated as hell compared to opensource desktop composite managers it is still very light on the GPU compared to games. Only thing that might slow it down is less memory bandwidth but all IGP's have at most as much bandwidth as CPU. Intel with its external memory controller is actually in a better position in that regard compared to AMD.

Scientia from AMDZone said...

enumae
Maybe I should have worded it differently, but what I said is still correct in the overall sense.

I should have said AMD won't be at total capcaity until the end of 2008.


And, that would still be incorrect. FAB 36 will be at full capacity by mid 2008 with 25K wspm. FAB 38 won't be at full capacity until the end of 2009 with 20K 300mm wspm. By end of 2008, AMD will be at roughly 78% total capacity.

Now after looking up SSE4 it would seem that K8L's new instructions are just a drop in the bucket.

That bucket seems to be missing a few drops. LZCNT isn't listed, perhaps because it is the Itanium killer instruction.

Does making your own chipsets at your own FAB save you money?

Yes, and no. If you use an older FAB to make chipsets then it saves you money. If you use a fully modern FAB which is capable of producing 65nm chips then you lose some of the savings. In this case, production of chipsets for Intel is a damage control option. It would be more expensive to allow the FAB capacity to set idle. Overall, I would still guess that Intel's production is a bit cheaper than using a Foundry but this advantage is more than lost by the northbridge.

Well I don't know what to tell you, but they have been doing it, and are aparently making money doing so.

What is your evidence that they are making money? How much?

Scientia from AMDZone said...

thekhalif
while RD690 and RD790 will provide, I believe, the basis for the Barcelona "glueless" chipset.


I don't know what you mean by "glueless". Any socket F chipset should work with nothing more than a BIOS update.

That is the next news I expect to hear from AMD, that they have updated 8000/8131 for 8Way+ SMP.

Not until 2008 with DC 2.0.

Ho Ho said...

scientia
"LZCNT isn't listed, perhaps because it is the Itanium killer instruction"

Could you describe what that instruction does and in what situations might it benefit? From what I know about this instruction, it has little to no effect.

Scientia from AMDZone said...

red
Have you not heard of Bearlake?


Yes, I've heard of Bearlake. Was there a point hiding in there somewhere?

Again, I fail to see the significance of graphics for bottom-of-the-barrel PC's.

And, you fail to see the significance of corporate sales and that this segment is growing faster for AMD than consumer sales. Since you missed it, the significance is money.

And it doesn't look like AMD is pushing their own platform very hard,

Well, that is probably because AMD doesn't have a platform. Did you miss that too?

I guess because they have no wireless/motherboards.

AMD doesn't have any motherboards; those are made by third parties. You seemed to miss this as well. However, AMD endorses both Broadcom and Atheros wireless components.

But yay stickers.

Did you miss this also? Dell, HP, Gateway, Acer, Lenovo and NEC have already signed up to receive their stickers. It appears they value the AMD stickers the same as the Centrino logo.

Red, was there anything you didn't miss?

Scientia from AMDZone said...

ho ho
Could you describe what that instruction does and in what situations might it benefit? From what I know about this instruction, it has little to no effect.


Well, come to think of it they did include POPCNT which would probably be used more often. As far as I know, LZCNT is the inverse of POPCNT. POPCNT counts ones and LZCNT counts zeros. Including POPCNT is definitely a blow to Itanium.

Ho Ho said...

Scientia
"POPCNT counts ones and LZCNT counts zeros."

Perhaps it's just me but in the last nine years of programming I have never thought I might need a functoin like that. Before you say it's because I program in high-level languages I'll say I have >5y of C/C++ experience.

Scientia
"Including POPCNT is definitely a blow to Itanium."

How come? Could you describe some algorithms that use that function to greatly increase its speed? Just saying that "it would speed up programs that need to count bits" does not qualify as an example :)

Ho Ho said...

I forgot to mention that Itanium has no LZCNT function, at least it is not listed in the manuals and instruction set references.

Red said...

Bearlake, in response to Erlindo. Chill off with the hostility?

Erm.. what does corporate sales have to do with graphics for bottom-of-the-barrel PC's being pointless? You're not insisting that corps play Solitaire 3D Extreme, are you?

Well, that is probably because AMD doesn't have a platform. Did you miss that too?
Yeah, but wasn't that the point of buying ATI?

AMD doesn't have any motherboards; those are made by third parties. You seemed to miss this as well. However, AMD endorses both Broadcom and Atheros wireless components.

Yeah, I already said that they didn't have motherboards, thanks for repeating me.
And that's why they can't do stuff like..
http://news.com.com/Intel+inside+again+for+new+Google+servers/2100-1014_3-6153431.html
http://news.com.com/Intel+shows+pint-size+server+motherboard/2100-1010_3-6120352.html

What is your evidence that they are making money? How much?
Well where is your evidence that they are losing money?

enumae said...

Scientia
And, that would still be incorrect.

For some reason the white background is not extended to the edge of the text, and the number feel on that edge, yes you would then be correct, sorry.

That bucket seems to be missing a few drops. LZCNT isn't listed, perhaps because it is the Itanium killer instruction.

Wow, I wasn't trying to be mean, but to list one that isn't there is a pretty weak argument.

If you use an older FAB to make chipsets then it saves you money.

What chipset that Intel has is built using, as you say, a modern FAB?

What is your evidence that they are making money? How much?

How about the higher ASP. I am no economist, but if they have higher ASP than AMD in the server area while pricing Cloverton about the same as Woodcrest, I am gonna have to believe people would buy Cloverton. Wouldn't you?

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

Erlindo
Intel's IGP is no match for gaming and I really doubt that it will do fine in Vista (I'll take intel's word with a grain of salt on this one).

I think Ho Ho answered everything very well :).

Greg said...

Enumae, just because you have a high ASP, does not mean you have high enough margins to overcome administrative and corporate costs as well as non-chip production based industrial costs.

I anything, Intel having motherboard production facilities is a bad thing, because they obviously don't sell large amounts of them in the channel, no PC manufacturer uses Intel motherboards, and they're probably fairly expensive to operate.

Not only that, but AMD has the bulldozer platform, due out some time this quarter, but that's obviously not out yet, since they didn't have stickers until now (and you can't have a platform without STICKERS!!). Anyway, the fact that they're selling those stickers already is a really good sign, as there was some fairly delayed reaction to Viiv, as the PC industry seemed to be going through a chill effect in terms of platform acceptance.

Erlindo said...

Ho Ho wrote: When?
expect them this same year. I don't know the month exactly, but some sources say that it would be in the second quarter. I'll get the info and post it right here.

Ho Ho wrote: I don't know if average Joes are stupider in US but at least in Estonia shopkeepers tell their customers that when you want to play games, any IGP is slower than cheap dedicated GPU. I can't remember when I saw last PC withouht dedicated CPU. Btw, even at the company where we work we have dedicated GPUs even though we could happily live without them.

Please, not the whole world is perfect :D
Asian and south american PC users buy their computers with everything on mind. They don't upgrade their computers the way you and me do it. Instead, they buy cheap computers with integrated video that does the job well for them in many ways.
Jus as a little example, I go to a community college and in one of the computer halls we have 30 brand new HP machines with integrated video (I believe its Nvidias 6150 IGP) and believe me when I tell you that some of my friends (including myself) improvise a little lan party in the hall and we play games such as Battle Field 2 and counter strike. This freat would be almost difficult (to not say impossible) in an intel platform thanks to their crappy IGP. ;)

HO HO wrote: Want to bet that won't happen before Intel stops creating IGP's?
It's just a matter of time when the average joe and corporate user see the great advantage of using an AMD platform. AMDs Better by Design will be competing with intel's vpro and Centrino platforms.

Ho Ho wrote: Because ...?
Even though Aeroglass is bloated as hell compared to opensource desktop composite managers it is still very light on the GPU compared to games. Only thing that might slow it down is less memory bandwidth but all IGP's have at most as much bandwidth as CPU. Intel with its external memory controller is actually in a better position in that regard compared to AMD.

I still think that intel's IGPs will suck in Vista (in the near term). :D

enumae said...

greg
...because they obviously don't sell large amounts of them in the channel, no PC manufacturer uses Intel motherboards, and they're probably fairly expensive to operate.

Just want to point out that my discussion with Scientia was in regards to chipsets, not motherboards.

To be clear, Intel does sell platforms, and they sell them to PC manufacturers.

A platform from Intel would include a motherboard, chipset and a processor.

This is why AMD bought ATI, to be able to provide a platform to PC manufacturers, they do not have the FAB's to produce motherboards but they will have the motherboards made.

Maybe I do not understand your point, but I believe you are mistaken.

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

In regards to margins, I do not have the vocabulary to explain my position clearly.

Here is what I see hapening...

I believe that Intels size and future FAB's will allow them to continue to set the price for Desktop, Server, and Mobile processors and with that ability comes the ability to control AMD's margins, which could halt AMD's future abilities to expand.

enumae said...

erlindo
I still think that intel's IGPs will suck in Vista (in the near term). :D

You need to look at what people who have actually used Intels IGP for Vista say about it, and no I am not one of them.

Do you believe that a company buying 200, 500 or 1,000 laptops cares about a slight advantage in 3D and at what cost?

TheKhalif said...

while RD690 and RD790 will provide, I believe, the basis for the Barcelona "glueless" chipset.

I don't know what you mean by "glueless". Any socket F chipset should work with nothing more than a BIOS update.

That is the next news I expect to hear from AMD, that they have updated 8000/8131 for 8Way+ SMP.

Not until 2008 with DC 2.0.



AMD can't wait until then to update 8000/8131. Just as they need their own chipsets for desktop, they need one for server also. Barcelona is too important to have ONLY a 3rd party chipset, IMHO.


I mean think about it. nForce Pro only supports 667DDR2. Barcelona supports at least 800, maybe even 1066 with the Shanghai refresh.

This way they can sample everything from CPU-chipset-StreamProc-IGP.

Scientia from AMDZone said...

red
Yeah, but wasn't that the point of buying ATI?


No. AMD doesn't do platforms like Intel; you should know this by now.

Yeah, I already said that they didn't have motherboards, thanks for repeating me.

No. You said that AMD didn't have wireless motherboards. If you were attempting to state that AMD didn't have any motherboards then it was unnecessary to qualify it by limiting your statement to wireless. I didn't repeat you; I corrected you.

Scientia from AMDZone said...

ho ho

You may be right; Intel may not consider LZCNT to be as important.

Adding POPCNT to an X86 platform is a blow to Itanium because previously Itanium destroyed its competition on certain benchmarks by using this instruction. I believe this code is used for things like counting parity bits. This would then have application in encryption/decryption, some types of software driven I/O, some types of data redundancy, and other applications. These would be more likely to be used in database applications than something like graphics processing.

Red said...

Enumae, just because you have a high ASP, does not mean you have high enough margins to overcome administrative and corporate costs as well as non-chip production based industrial costs.
So you are also suggesting that Intel is losing money on these;) So elaborate on the administrative, corporate, non-chip production base industrial costs. Whose margins rose last quarter again?

I anything, Intel having motherboard production facilities is a bad thing, because they obviously don't sell large amounts of them in the channel, no PC manufacturer uses Intel motherboards, and they're probably fairly expensive to operate.
They are making $$$ off of them, as opposed to Flash, so they are a good thing.
http://accessories.us.dell.com/sna/productdetail.aspx?c=us&l=en&cs=04&sku=A0935523
http://h10025.www1.hp.com/ewfrf/wc/genericDocument?docname=bph01186&cc=us&lc=en&dlc=en&dlc=en&lang=en
http://support.gateway.com/support/supinfo/index.asp?pg=2&file=dt_mot.html
And just wondering, if not the channel, and according to you, not OEM's, I guess they are just making these for fun. Fairly "expensive to operate"? Running out of talking points? xD

Bulldozer is the mobile chip.
http://www.tgdaily.com/2006/08/30/amd_ati_yokohama/

I said wireless/motherboard, not wireless motherboard. Sorry that you don't understand the subtleties of punctuation.

No. AMD doesn't do platforms like Intel; you should know this by now.
ATI likewise gives AMD an all in one solution for OEM's who prefer this to putting together systems piecemeal.
Umm.. Yeah, so with buying ATI, all that they have to show for it is some "open" platforms, with pretty stickers with components already used in the past anyways.

Scientia from AMDZone said...

enumae
believe that Intels size and future FAB's will allow them to continue to set the price for Desktop, Server, and Mobile processors and with that ability comes the ability to control AMD's margins, which could halt AMD's future abilities to expand.


No. This worked back in 2002. It does not work today. Because of AMD's second FAB and AMD's movement into both server and mobile areas, Intel can no longer set these prices without hurting itself more. Intel cannot control AMD's margins.

Scientia from AMDZone said...

thekhalif
AMD can't wait until then to update 8000/8131. Just as they need their own chipsets for desktop, they need one for server also. Barcelona is too important to have ONLY a 3rd party chipset, IMHO.


I wasn't talking about chipsets. No opteron will be capable of glueless > 8-way until DC 2.0 in 2008. I don't know what chipsets will be available before then. However, greater than 8-way is not a function of the chipset.

enumae said...

Scientia
Because of AMD's second FAB and AMD's movement into both server and mobile areas, Intel can no longer set these prices without hurting itself more.

I understand that Intels margins were down for 2006, and that AMD gained market share at least thru Q3 and possibly Q4.

But here is an example (theoretical) of what I see...

100 chips (65nm).

Intel sells 70 at $45
AMD sells 30 at $45

Each chip cost $20

Intel makes $1750
AMD makes $750

Intel cuts the price down to $40, now consider these chips are equal in performance terms.

Intel sells 70 chips at $40
AMD sells 30 at $45

Intel makes $1400
AMD makes $750

Now factor in 45nm, and a reduction in chip cost to say $15 while AMD is still at $20.

Intel sells 70 chips at $40
AMD sells 30 at $45

Intel makes $1750
AMD makes $750

Intel can charge less and make the same amount due to there manufacturing advantage, thus regain market share due to a lower price.

Know I may not be making it very clear and this is theoretical, but this should make it pretty simple to follow.

Scientia from AMDZone said...

red
I said wireless/motherboard, not wireless motherboard. Sorry that you don't understand the subtleties of punctuation.


Actually, I do. You cannot use the forward slash to join two dissimilar items. So, if you actually meant "wireless components or motherboards" you should have written it that way. You are again corrected.

Umm.. Yeah, so with buying ATI, all that they have to show for it is some "open" platforms, with pretty stickers with components already used in the past anyways.

Are you still failing to grasp the concept of factory endorsement? AMD has been doing this for awhile now with motherboards. Can't you understand how this could be applied to other components? There is also the concept of competitive merchandizing.

However, since you are so derisive then try explaining how Centrino as a platform is better than an AMD endorsed platform bundle. Exactly what is it that you get from Centrino that you cannot get from an AMD endorsed system? I can't wait to hear this.

Red said...

Centrino is not better tech wise. Centrino just happened to become synonymous with wireless computing. What will "Better by Design" be synonymous with?

AMD endorsed? It'll only take up 1/3 to 1/2 of the sticker xD

Please explain to me, what the point of "Better by Design" is, that a simple "AMD is Awesome" sticker can't do. I can imagine that it would confuse customers, between Nvidia vs ATI, Broadcom vs Atheros, as if Intel vs AMD isn't confusing enough.

All-in-one:ATI likewise gives AMD an all in one solution for OEM's who prefer this to putting together systems piecemeal.
Factory endorsement:No. AMD doesn't do platforms like Intel; you should know this by now.
All-in-one solution or factory endorsement of 3rd parties. Pick a side and stick with it.

Scientia from AMDZone said...

enumae
Intel can charge less and make the same amount due to there manufacturing advantage, thus regain market share due to a lower price.


Well, this would mean that Intel would be reducing its ASP and margin. However, in Q4 it appears that it was AMD that did this. It gets harder for Intel to do this during 2007 because AMD's costs will drop more quickly. Intel will get a window of about two quarters to gain advantage from 45nm.

Fujiyama said...

Scienta can you explain more about Direct Connect 2.0. I was trying to find any info about that. I always have on mind that AMD is going to split the chip architecture to design or even produce the microprocessor in a smaller parts. This is what Hester was saying about modular design when you take only these microprocessor parts what you need to create a final product. So if you need faster GPU or FPU or larger cache - you produce some parts and connect them using HT links or misterious DC 2.0. Can you explain that?

Red said...

You still aren't making any sense. You say that now that they've bought ATI so that..
ATI likewise gives AMD an all in one solution for OEM's who prefer this to putting together systems piecemeal.
Yet now say that OEM's will love AMD endorsed per vendor stickers. Well they could've saved $5B by just endorsing them before, so what exactly does ATI add to them again?

If you still aren't getting this then this is about choice, something that Intel does not offer.
Or, a simple "AMD is Awesome" sticker, with vendors that met those requirements would've sufficed, just like the Centrino, Viiv, or Live! stickers. Why exactly do they need to differentiate the wireless vendor, except to remind everyone that they are also wireless?

enumae said...

Scientia
Well, this would mean that Intel would be reducing its ASP and margin.

But at the expense of AMD, look at what Conroe did for desktop prices, and the effect it had on AMD's margins.

Also look what Intel did to AMD's Servers ASP.

However, in Q4 it appears that it was AMD that did this.

As a direct result of Intel's pricing.

AMD is not in a position to dictate the prices, not since the launch of Conroe, Woodcrest and Cloverton.

The only window AMD will have ths year to dictate ASP will be between Q3 and Q4, K8L in volume and Intel introducing 45nm.

It gets harder for Intel to do this during 2007 because AMD's costs will drop more quickly.

I don't think so.

Who is larger and who is already largely at 65nm which is more cost effective than 90nm?

Do you think Intel is worried about tha margins at 90nm or 65nm more?

Intel will be ramping Core architecture so that by the end of Q2 Netburst will be about 10%.

(I read that somewhere, I can find the link if you want it).

Will AMD be completely transitioned to 65nm by the end of Q2?

Intel will get a window of about two quarters to gain advantage from 45nm.

Intel will get the benefit for much longer than two quarters in Mobile and Desktops.

The two quarters you are talking about would primarily be aimed at the Server area, I think that will be Q1 and Q2 2008.

TheKhalif said...

I wasn't talking about chipsets. No opteron will be capable of glueless > 8-way until DC 2.0 in 2008. I don't know what chipsets will be available before then. However, greater than 8-way is not a function of the chipset.


That is what the L3 is for. The chipset itself still needs to come from AMD. They are saying that Barcelona will hold up to 32 sockets because of link unganging.

I realize DC2.0 is coming but Barcelona is an important platform. I will be surprised if they don't release a server chipset.

They are releasing RD790 for the desktop Stars chips.

TheKhalif said...


Intel will be ramping Core architecture so that by the end of Q2 Netburst will be about 10%.

(I read that somewhere, I can find the link if you want it).

Will AMD be completely transitioned to 65nm by the end of Q2?

Intel will get a window of about two quarters to gain advantage from 45nm.

Intel will get the benefit for much longer than two quarters in Mobile and Desktops.

The two quarters you are talking about would primarily be aimed at the Server area, I think that will be Q1 and Q2 2008.





I think you read that wrong. Core 2 is not due to even be 50% until sometime this quarter.

Why can't you guys just accept the duopoly? It's not going anywhere.

Scientia from AMDZone said...

Red

I'll try one more time to explain this to you.

No one who buys a Centrino wireless notebook cares whether or not all of the components came from Intel. The only one who really cares about that is Intel because they make more money by selling all of the components.

However, vendors are not happy about that because it removes their flexibility and it also creates a brand that is stronger than their own. HP wants you to identify with HP first, not the Centrino brand.

AMD provides a merchandizing vehicle (the stickers) that allows for flexibility and does not overshadow the OEM brand. However, this is more than just a sticker because the configuration is endorsed by AMD. This ensures quality instead of just letting anyone with an AMD notebook slap a sticker on the case.

The purchase of ATI provides a cpu, chipset, and graphics solution for those OEMs who want this. This should get more customers for AMD. However, not all customers are looking for cpu, chipset, and graphics from one vendor and they have the flexibility to configure as they wish.

I know you keep trying to make the point that AMD does not build its own wireless connector. This is a trivial point since two other vendors do provide good solutions for AMD products. I'm writing this on a wireless AMD based notebook right now and the fact that the wireless connector is not made by AMD has not been a problem. The vast majority of customers will not care that AMD doesn't make the wireless connector. So, I don't understand your point. The most important thing is being able to put together a system that works properly and you can do that with AMD.

AMD's purchase of ATI will help ensure that AMD has good low power mobile chipsets in the future along with other needed products when it needs them. The strategic advantage of being able to plan its own chipset and graphics components is obvious to me; I don't know why it isn't obvious to you.

Scientia from AMDZone said...

Red

Okay, you asked why AMD is putting the wireless manufacturer on the sticker. For two reasons.

With component endorsements the items have to be tested with AMD hardware to make sure that they work. There is a cost involved. Sometimes the cost is absorbed by one party but typically it is paid by both. So, in this case Broadcom would share the cost of testing with AMD. When testing was finished AMD would endorse Broadcom with its hardware. The endorsement in this case is Broadcom's name on a sticker. This is the most direct way to show that AMD endorses Broadcom.

The second reason is customer preference. There may be customers who prefer Broadcom wireless to other manufacturers. So, this puts the wireless supplier's name where it is visible. To some extent this can also lend credibility to a wireless solution since Broadcom is very well known as a networking component manufacturer. In this case AMD would be seeking to gain credibility by association with Broadcom.

Scientia from AMDZone said...

the khalif & fujiyama

Okay, let me correct something I said earlier. HT 3.0 is capable of a split or unganged mode which would indeed allow greater than 8-way. Unganged mode essentially splits the HT link in half so one link can be used for two different connects although each at half the speed. However, since HT 3.0 is about twice as fast as HT 2.0 this would probably be reasonable.

Direct Connect architecture 2.0 increases the number of HT links to 4. This would get rid of the current problem of increasing latency when going higher than 4-way.

HT 3.0 was not originally due to be released until DC 2.0 however it may be that AMD is feeling pressure in servers from Intel and will release HT 3.0 early. AMD may release an Opteron with HT 3.0 in late 2007 so that would allow greater than 8-way before 2008.

Red said...

Thanks for deleting the inflammatory comment, I hope that you will not stoop that low (once) again.

Scientia, if Centrino buyers don't care about whether the components are Intel or not, why should "Better by Design" buyers care who makes the CPU/GPU/wireless? You're not making it clear how the multi-vendor sticker is any more better than a single sticker, given to those that meet the sticker's requirements. I think a single "AMD is Awesome" sticker/platform would be more able to compete with Centrino, as opposed to "Better by Design" competing with itself, don't you?

If you're going to say that not having wireless is a trivial point, you could say the same thing about AMD not having chipsets before, they could've saved $5B and just endorsed it like the wireless. I'm not saying that having wireless is a big deal, but it seems to me that you're contradicting yourself.

The strategic advantage of being able to plan its own chipset and graphics components is obvious to me
I actually don't get this. What does GPU have to do with CPU? And IGP and chipset are basically together, so I don't see the need to differentiate them. Please elaborate.

Red said...

"Customer" preference? Are you referring to consumers or OEM's? The difference in wireless vendors is like the differences in salt makers (wireless is wireless, make it cheap and make it work).

Scientia from AMDZone said...

enumae

First of all, Intel did not push down AMD ASP's in servers, nor in mobile, just on the desktop. However, desktop is the highest volume so this had some effect.

AMD's costs will drop faster than Intel's during 2007. Intel's only cost saving is by increasing Conroe production which has a smaller die than P4. That is it until 45nm. AMD on the other hand steadily gains cost reduction as 300mm becomes a higher percentage of production. Likewise AMD gains a reduction in cost as it shifts more to 65nm.

Again, as AMD's costs come down Intel will be unable to control price. About the only remaining advantage that Intel has is volume. Intel still has the capacity to gain volume share if the market increases rapidly enough. However, this is a grim prospect for Intel since AMD's volume increases everyday and by the end of 2009 AMD will be fielding two full 300mm FABs. This means that Intel's previous volume leverage will be reduced by a factor of about 3. It will be impossible for Intel to maintain a standards monopoly under those circumstances. AMD is hoping that this will occur earlier by the end of 2008.

Scientia from AMDZone said...

Well, as a customer, the OEM would not care about the sticker. The sticker is obviously for the buyer.

However, the OEM does care about the endorsement by AMD because that should guarantee that it will work properly. Cost is also there too since OEM's have the choice of two wireless vendors instead of just one.

Red said...

First of all, Intel did not push down AMD ASP's in servers, nor in mobile, just on the desktop. However, desktop is the highest volume so this had some effect.I've already covered this in an earlier comment but you seem to have disregarded it.

http://seekingalpha.com/article/18742
Gross margin was 51.4%, down from 56.8% in the second quarter of 2006, largely due to lower desktop ASPs, as prices declined faster than cost reductions, primarily in the desktop area.
Server and mobile ASPs were up, and desktop ASPs were down sequentially.


http://seekingalpha.com/article/24929
The decrease from the prior quarter was largely due to significantly lower server processor ASPs, and the inclusion of ATI operations.
Server processor unit shipments were essentially flat quarter over quarter, and ASPs declined significantly
Well, the mobile ASP was slightly up and the desktop ASP was slightly down, but marginal.

Therefore, desktop has hurt AMD in Q3, servers in Q4, though I don't see much new on Intel's side to be hurting AMD's mobile biz.

AMD's cost reductions will only make up for their drop in gross margins. They don't seem sure of themselves that they can bring themselves back to 50%.

http://www.dailytech.com/article.aspx?newsid=5414
What exactly does "control" pricing mean? As in AMD having to follow suit after Intel cuts prices? Hmm.. E4300 for $113, I smell sub $100 X2.

Red said...

1. No one who buys a Centrino wireless notebook cares whether or not all of the components came from Intel.

2. The sticker is obviously for the buyer.
Right.. So again, what's the point of specifying the vendors of the sticker?

I still see this "Better by Design" scheme as confusing for consumers and competing within itself.

Scientia from AMDZone said...

red

Yes, I stand corrected on the server ASP's; obviously they did drop in Q4.

What exactly does "control" pricing mean?

Where did you see this term? I didn't see it in the TomsDaily article.

It doesn't compete with itself because the offerings are determined by the OEM's. This is simply secondary branding within an OEM's product line which is what they prefer.

Red said...

Intel cannot control AMD's margins.

Again, as AMD's costs come down Intel will be unable to control price.

I see no TG link referenced, besides my own correcting Greg on Bulldozer.. So what does control pricing mean? By AMD not being "controlled", are you suggesting that they will let the E4300 stay comfy in its future spot as price/perf dual core king (assuming that AMD doesn't want to be "controlled").


Competing within themselves on the market place. X2 is the nonvariable, but ATI GPU+Atheros, ATI GPU+Broadcom, ATI IGP+Atheros, ATI IGP+Broadcom, Nvidia GPU+Atheros, Nvidia GPU+Broadcom, Nvidia IGP+Atheros, Nvidia IGP+Broadcom. Centrino is just a pretty sticker, AMD is making things complicated. If one cared about notebook specs, they would look beyond vendor stickers anyhow. There is also nothing about the sticker that accentuates the AMDness or the "platformness".

enumae said...

Scientia
First of all, Intel did not push down AMD ASP's in servers, nor in mobile, just on the desktop. However, desktop is the highest volume so this had some effect.

I did not mention Mobile, but I do appreciate your abiility to acknowledge when you are wrong as you did in a response to Red's comments.

AMD's costs will drop faster than Intel's during 2007...

I see your point. Now what effect will 45nm have on that transition considering AMD won't be at total capacity for FAB 36 at 65nm until the middle of 2008? Also keep in mind that while AMD is ramping so is Intel, but at 45nm.

...However, this is a grim prospect for Intel since AMD's volume increases everyday and by the end of 2009 AMD will be fielding two full 300mm FABs. This means that Intel's previous volume leverage will be reduced by a factor of about 3....

First I am not hoping for Intel to have a monopoly but, I do not believe that you are factoring in the additional cores that will be (presumably) the norm in 2009. Two fabs at 300mm is great for dual cores, but not Quad cores.

Again, as the number of cores increases and the transition to 45nm the smaller the effect of AMD's two 300mm FAB's will have. Everything I have read is pointing to Intel moving very quickly to multiple core's (quad cores).

An example is the number of WSPM at the new fab in Israel, 30,000 WSPM on 300mm wafers at 45nm in the middle of 2009. This is only one 45nm FAB, there are two others that should comee online either later this year or early 2008.

From what I have shown you, you have to admit that AMD's two FAB's will be very hard pressed to apply any pressure on Intel while the Server and Desktop move to Quad core over the coming years, all the while also accepting the fact that if Intel is able to do what they are planning to do, they will also have the ability to control ASP's in both Server and Desktop segments.

enumae said...

I need to make a correction.

I had previously stated...

This is only one 45nm FAB, there are two others that should comee online either later this year or early 2008.

I should have made it a little clearer, the Oregon FAB is already built, but the Arizona FAB is/will be new.

Also, someone at [H]ard Forum had posted a link to an interesting wesite I had not heard of, never the less there are some interesting interviews/tours (in video) with some of Intels people.

Here is a link to the Intel video section.

Greg said...

Red, if you saw the pictures of the stickers, you'd notice there were only three, and that they all basically looked the same.

As to why the customer would care, they really wouldn't, however, most people who care, don't like the fact that Intel makes their wireless chipsets, because they lack competition and thus suck. Especially if you use linux. If you remember the early days of centrino, the wireless chip was often limited to 1Mb/s even though most chipsets out there supported 11Mb/s. I still have a few banged up laptops at home that evidence of this. With AMD's approach, you have an option, if you care, and you have competition between brands for higher speeds. Chances are, that those branding stickers will eventually show the connection speed.

Another problem that Intel will face is that right now they're trying to force Wireless N by including it in their next platform, even though the specification wont even be finished by then, much less will any vendor supply it. This could backfire if they try to use pre-N chipsets, and incompatibilities with N routers become a hallmark of the brand.

My question is, how much does it cost to ramp an entirely new, never before used fab or two others like it, compared to ramping a fab has already been used. This could be a disadvantage for Intel as AMD begins ramping 300mm at fab30. As I've said before, regardless of the fact that they end up using them for other purposes, it can't be more economically sound to build a new fab when you want to move to a new process than it is to upgrade an already existing fab.

How can 300mm be better for dual core, but not better for quad core? The extra wafer space can be used for anything, regardless of how many cores. Obviously, this is all comparative to what AMD had before, and as long as it's better, it's bad for Intel. 2 fabs> 1, and although barcelona is their largest core yet, they're not going to be producing only that processor, and if it's faster than Intel's quad core processor, they can demand a price premium that justifies that.

enumae said...

Greg
My question is, how much does it cost to ramp an entirely new, never before used fab or two others like it, compared to ramping a fab has already been used.

New - Israel Fab (Fab 28)
$3.5 billion US dollars

New - Arizona Fab (Fab 32)
$3 billion US dollars

Conversion - Arizona (Fab 12)
$2 billion US dollars

This could be a disadvantage for Intel as AMD begins ramping 300mm at fab30. As I've said before, regardless of the fact that they end up using them for other purposes, it can't be more economically sound to build a new fab when you want to move to a new process than it is to upgrade an already existing fab.

It may not be more economical, but as the market grows and as you add more cores you need FABs' in order to keep up with demand, also keep in mind that(I may be off slightly) Intel makes a chipset for about 70-80% of its processors. Thats alot of chipsets and at a larger node you would need more capacity.

How can 300mm be better for dual core, but not better for quad core?

My comment is related to quad cores, in the sense that as the market shifts away from Dual core to Quad cores (Server and Desktop) they will once again be capacity constrained.

In the Q4 confrence call they stated that they were only at a 30% Dual core mix, and yet they were still capacity constrained.

What do you think Quad core will do to there capacity issues?

The extra wafer space can be used for anything, regardless of how many cores. Obviously, this is all comparative to what AMD had before, and as long as it's better, it's bad for Intel.

I understand that, but this is why I keep trying to get my point accross that Intel, by looking at what I have read, seems to be moving the market to Dual and then Quad cores and in doing so will again hamper AMD's achivements in market share gains and revenue gains.

2 fabs> 1, and although barcelona is their largest core yet, they're not going to be producing only that processor, and if it's faster than Intel's quad core processor, they can demand a price premium that justifies that.

AMD's claims have come about as their Server ASP is declining.

They have not mentioned any specific test, there comment is aimed at Floating Point performancec, an area in which they are already ahead of Intel on the Server front, probably some 15-20%.

It sounds like a company trying to keep people from going to Intel, and without a solid release date, a company that knows it has hard times ahead.

As you said their largest core, which will lower capcity even more, and charging a premium to justify the size will only bring there growth to a halt.

I understand I am vocabularily constrained :), but I see my point and I hope I am making it clear to you.

Red said...

As to why the customer would care, they really wouldn't

Exactly

however, most people who care, don't like the fact that Intel makes their wireless chipsets, because they lack competition and thus suck.

And if they did care, they wouldn't base their choice of a laptop because of their favorite wireless vendor (last time I checked, wireless is wireless) is on a sticker.

I can imagine, Grandma watches TV, she sees an orange/red/blue sticker with Kelly Ripa, goes to PC store, demands that sticker, but the store only has orange/green/red stickers, and confuses her. Or, Grandma sees Centrino on TV, wants a Centrino, and gets one.

So. It has been noted that people don't care about the vendors, and if they did, they would take note of the specs, not the sticker. And how is a multi-vendor sticker better than a single sticker again?

On 802.11n..
http://www.80211n.com/intensi-fi-products.html
http://www.apple.com/wireless/80211/
http://www.atheros.com/news/xspan.html
http://en.wikipedia.org/wiki/IEEE_802.11#Pre-n_equipment

Scientia from AMDZone said...

enumae

Quad on 45nm is about the same as dual core on 65nm and single core on 90nm. AMD will absorb the core increases with process size.

However, there is no denying the increase in capacity for AMD. Intel's position today is much worse than it was from 2002 - 2005 when AMD only had FAB 30. AMD will increase from a capacity that can handle 10% of the market volume to a capacity that can handle about 33% of the market volume. This is a geniune loss for Intel. Recall too that Intel scaled back its original construction plans. Intel is giving ground to AMD.

As far as the performance of K8L goes, it will easily beat any Intel Xeon running with FBDIMM. With both systems running with registered DDR2, dual core should be about the same however AMD should stay slightly in the lead with quad core. This would for single and dual socket systems. Intel should be competitive with 4-way dual core but less so with 4-way quad core. And, Intel will fall behind for greater than 4-way. This is where 45nm may be able to help Intel be competitive for 4-way. However, Intel then gets hammered by AMD's 45nm, DC 2.0, and gpu processing. My guess is that at that point Intel will start paper launching its new socket architecture which won't be released until 2009. I'm not sure what will happen in 2009 because AMD should be doing segmented core updates by then. Basically, I see a lot of back and forth between AMD and Intel between now and 2009. I don't believe that Intel will be able to stay in the lead during this timeframe.

Red

I was saying that AMD's greater cost reductions during 2007 will prevent Intel from squeezing AMD on price. In other words, Intel will not be able to price its chips so that it has good margins while AMD does not. This was a common strategy for Intel through 2002 and 2003. It was less effective in 2006 and will be less effective in 2007.

I'm sorry but your sticker comparison is bogus. HP for example puts a lot more information on a notebook than just the cpu vendor's sticker. Essentially what you are trying to do is ignore 80% of the stickers on the notebook and try to make the argument about only those from AMD and Intel. In your argument, Grandma would see the Centrino sticker and then be overwhelmed by all the other information on stickers on the notebook. This is nonsense.

AMD's server ASP's should come back up by Q2 both due to additional 65nm production and the introduction of K8L.

Intel does gain an advantage for perhaps 6 months when it introduces 45nm. However, half of this timeframe overlaps with the 300mm startup of FAB 38 and the still increasing capacity of FAB 36. So, while it will have some effect it will be less than it would otherwise. AMD will probably also be demonstrating gpu processing during this time so this too will tend to dull Intel's advantage.

Roborat, Ph. D. said...

Scientia wrote: "This loss of contracts created a big loss of revenue for ATI in the fourth quarter which again was expected"

Incorrect. ATI's contract only expired the end of the year. ATI's has been on a downward trend in terms of profit since early 2006 and has only been purely losing out to NVIDIA. The impact of losing Intel's market has yet to have an impact.
Long term wise, a scaled down ATI will turn a profit. The margins though will have a more permanent affect on AMD.


Scientia wrote:
"They will introduce K8L Barcelona in Q2 which should make 2007 K8L's year in the same way that 2006 was for C2D"
Don't you think that that is a pretty bold statement? AMD's vague description of K8L performance advantage over the competition is very alarming. Key note they mentioned is "performance per watt". And we all know the clock frequecies are going to be very low while Intel can easily modify the clock to get "best performance" or best performance/watt" depending on what they want.
K8L needs to trash Woodcrest/Clovertown to become like "C2D". I seriously doubt that. Best case, AMD comes within parity. The problem is Q4'07 45nm kicks in.

TheKhalif said...

First I am not hoping for Intel to have a monopoly but, I do not believe that you are factoring in the additional cores that will be (presumably) the norm in 2009. Two fabs at 300mm is great for dual cores, but not Quad cores.

Again, as the number of cores increases and the transition to 45nm the smaller the effect of AMD's two 300mm FAB's will have. Everything I have read is pointing to Intel moving very quickly to multiple core's (quad cores).



AMD has said they are NOT going to get into the "core" race. That is the reason for Fusion and Torrenza.

Barcelona's die is not that much bigger than X2 1MB. Because we are mainly looking at going from 200mm to 300mm wafers, there will still be a cost savings.

Scientia is right that Intel can no longer use their volume to push AMD around and if they work overtime to shift to 300mm at Fab30, then that means a clear path to installing 65nm equipment by Q3.

Then at that point they can rotate a section to 45nm equipment by Q4. That should allow revenue shipments by Q208.

I believe Fishkill already has 45nm test chips and equipment.

I just can't understand why people doubt that Barcelona will improve upon Opteron.

I have closely followed all of the info about the arch and AMD should definitely get 40% greater IPC (general; both int and fp).

Because in this case the dual core will also have L3, latency issues are basically erased and the same advances in Barcelona will be in Kuma.

From the way AMD described the increases, they either have Kuma samples or deactivated two cores to get the 1.8x/3.6x numbers.


I think we can safely say that the perf/watt will be much better than 90nm Opteron since everything has been widened and enlarged while fitting in the same envelope.


I bet AMD feels like Roger Maris when he said "why can't they have enough room in their hearts for two?"

Aguia said...

AMD's server ASP's should come back up by Q2 both due to additional 65nm production and the introduction of K8L.

Why do you think that Scientia?

-Because AMD can paper lunch K8L like Intel did with Core 2 Duo forcing Intel to lower prices?

-And if Intel doesn’t do it, charges a price premium for the new parts like they did with Athlon 64/X2 before Core 2 Duo arrived?

I know that Intel native quad core will only arrive in 2H 2008 leaving AMD with one full year in advance over Intel, but isn’t the Intel 45nm process remove the gap since Intel can manufacture cheaper* Double dual core processors?

*(and other benefits)

Ho Ho said...

thekhalif
"I have closely followed all of the info about the arch and AMD should definitely get 40% greater IPC (general; both int and fp)."

Could you please be more specific? Almost all the information I've seen so far suggests that almost every improvement they make is already there in C2D. Logically that would bring it to the same level as C2D. Benefit will come from having IMC and maybe a bit higher bandwidth per socket.

S said...

There seems to be a assumption in the discussion that AMD will have Barcelona out by Q2. Is AMD close enough to get Barcelona out in Q2 ? I know there were statements made by AMD senior folks about production wafers coming out soon. Shouldn't engineering samples have been out by now in such case ? Or at least samples to look at like Intel had with Penryn ?

TheKhalif said...

Could you please be more specific? Almost all the information I've seen so far suggests that almost every improvement they make is already there in C2D. Logically that would bring it to the same level as C2D. Benefit will come from having IMC and maybe a bit higher bandwidth per socket.

I am not basing that statement on anything to do with Core2. I am basing it on what the improvements per core over K8 SHOULD be.

Do I have to list all of the improvements for the 100th time?

Int improvements
256bit L1-L2
dual 128bit loads/cycle
enhanced prefetch
out-of-order loads (some stores)
32B fetch width(2x K8)with additional retire logic
smaller macro-ops


FP Improvements
2x 128bit FP 4 DP FLOPS/cycle
2x128bit SSE4A
dual loads/cycle
larger branch history(helps both)


All of these will add up to 40% per core especially when including L3.

But then, AMD engrs may not be good enough to improve upon K8 so I can see where you could have doubts.

Scientia from AMDZone said...

Red
All Sci does is bitch and moan. Write 6 paragraphs about something which he could probably get across in 6 lines.


You are done. Don't post here again.

TheKhalif said...

There seems to be a assumption in the discussion that AMD will have Barcelona out by Q2. Is AMD close enough to get Barcelona out in Q2 ? I know there were statements made by AMD senior folks about production wafers coming out soon. Shouldn't engineering samples have been out by now in such case ? Or at least samples to look at like Intel had with Penryn ?


Here is a pic.

http://www.watch.impress.co.jp/akiba/hotline/20070203/image/nvis36.jpg"

Scientia from AMDZone said...

roborat

If Intel's contracts continued until end of 2006 then I guess you would be right that the loss of revenue cannot be attributable to the merger. However, this would really only strengthen the argument for the merger since ATI will obviously benefit.

Yes, I'm certain that K8L will be released in Q2. However, these have already been bought up so there won't be any in the channel.

AMD does have engineering samples of Barcelona however AMD doesn't pass theirs out like candy as Intel does. AMD's ES's are notoriously hard to get ahold of however a few really determined people have gotten them.

I expect that Barcelona will still trail C2D at the same clock in integer performance but it should close most of the current gap. However, it should be very close in SSE performance whereas now K8 trails by a large margin.

Thekhalif

Yes, those improvements plus adding additional stack processing to remove stack instructions from the pipeline, better ordering of bus access plus splitting the two halves of the memory controller. Also some more fast path instructions.

It is possible that this could catch up to AMD in integer but I'm currently estimating this. I'm basically saying that this should remove about 2/3rds of the current gap.

Scientia from AMDZone said...

Red

I'm sure you'll have plenty of laughs with your friends at Forumz and Extreme Systems. Or, you can start your own blog. That way you could demonstrate how 6 paragraphs can be reduced to 6 lines. Also, you can state the real truth about Intel instead the ramblings of an AMD fanboy.

Goodbye.

TheKhalif said...

Yes, those improvements plus adding additional stack processing to remove stack instructions from the pipeline, better ordering of bus access plus splitting the two halves of the memory controller. Also some more fast path instructions.

It is possible that this could catch up to AMD in integer but I'm currently estimating this. I'm basically saying that this should remove about 2/3rds of the current gap.



I may just be a little "jaded" as I can honestly see the 1.8x dual core(they're quoting FP). Since Core 2 is only 15-20% faster per "dual core" AMD only needs to make Kuma 25-30% faster.

It's been a long time since I've seen numbers but I ntel managed to stay ahea of Athlon MP with L3, so I'm expecting (because of decreased latency) 10% just for the L3.

That lowers the necessity of the other architectural improvements to 15-20%.

I see that as entirely doable. Another major factor I see is that Barcelona is designed for 65nm and switching speed alone may increase by, according to AMD, 40%.

If we just count that as 10%(25% efficiency), the architectural enhancements only need to be 5-10%.

Brisbane is not a good candidate for analysis of 65nm as it is just a shrink of a 130nm design, IMHO.

It does say a lot for AMD's engr process though. I think they will continue to do what they say and execution should improve from H206 since they were obviously dealing with acquisition issues and of course a "price war."

Scientia from AMDZone said...

roborat
Don't you think that that is a pretty bold statement?


No. This will be the only new architecture introduced in 2007.

AMD's vague description of K8L performance advantage over the competition is very alarming.

Why? I don't think it will have any signficant performance advantage; I think it will mostly just erase Intel's.

Key note they mentioned is "performance per watt".

Which is probably the only real advantage or Intel.

And we all know the clock frequecies are going to be very low

There have been rumors. Not sure if that is the same as knowing.

while Intel can easily modify the clock to get "best performance" or best performance/watt" depending on what they want.

Really? What makes you think that?

The problem is Q4'07 45nm kicks in.

At low volume on both D1D and the brand new FAB 32. AMD will deliver 45nm mid 2008 along with DC 2.0 and GPU processing.

enumae said...

Scientia
Quad on 45nm is about the same as dual core on 65nm and single core on 90nm. AMD will absorb the core increases with process size.

I understand your point, but AMD does not have 45nm at there expense and will not have it until the end of Q2 2008 and not with any volume until the end of Q3 2008.

Looking at the presentation by:

Bob Rivet
Executive VP and Chief Financial Officer
December 14, 2006

Looking at the graph on page 17, you can see what percentage of wafer outs will be for 45nm, 65nm and 90nm. The numbers I am showing are for the end of each quarter.

Q4 2006
65nm = 3%
90nm = 97%

Q1 2007
65nm = 24%
90nm = 76%

Q2 2007
65nm = 41%
90nm = 59%

Q3 2007
65nm = 55%
90nm = 45%

Q4 2007
65nm = 67%
90nm = 33%

Q1 2008
65nm = 78%
90nm = 22%

Q2 2008
45nm = 3%
65nm = 81%
90nm = 16%

Q3 2008
45nm = 25%
65nm = 67%
90nm = 7%

Q4 2008
45nm = 33%
65nm = 67%

I see the number one way, you see them another, I think the only thing left to do is sit and wait.

I just want to be clear, if Intel wins back market share or more revenue share in Q1 or Q2 in a significant amount AMD will be in trouble, would you agree with that?

Scientia from AMDZone said...

enumae
The numbers I am showing are for the end of each quarter.


How would you know that? The graph doesn't say what the timeframe is. The marks could be beginning, middle, or end of quarter.

Wallachian said...

At low volume on both D1D and the brand new FAB 32. AMD will deliver 45nm mid 2008 along with DC 2.0 and GPU processing.

Coming from within the semiconductor industry, I can definitely say this:

AMD acquired ATI in Q4 of 2006, and started talking about their fusion plans. To plan out a new CPU and get all its development in place, they will need atleast 1 qtr or 2. Then usually it takes close to 2 years (if you have a good strong design team, maybe a qtr less) to complete the design of the product and tape it out. Then it takes a year for the backend silicon validation. Assumming that AMD+ATi's planning of the new CPU ends in q1'07, they will probably tape out the design in q4'08 (normally q1'09) and launch it in q4'09 (usually q1'10).

I doubt AMD can cook up the design, any quicker lest it risks a DOA product. Even if it has a modular CPU, a lot of changes will be needed. Particularly when you talk about that the new AMD cpu should also come with a next gen ATI gpu, and you have to plan the two architectures and protocols carefully. If someone wants to refute this, I believe in recent press reports, AMD has said it will deliver the product in mid-09, after earlier saying mid-2008. That mid-2008 was just a date to keep the analysts in a good mood at the time of the acquisition announcement.

Scientia from AMDZone said...

No, I'm talking about work in progress, not GPU on die. First silicon for this would have to be about mid 2007 and that is doable.

GPU on die, maybe late 2008 which would put first silicon late 2007.

BTW, your estimate is pretty good but GPU on die won't take that long. This is because the GPU design is basically just dropped onto the die and hooked up through the Crossbar. It is doable by late 2008.

Azary Omega said...

Scientia is right, they wont make a whole new design, what they will do is just add a GPU on different powerplane to the die. Also I'm not sure why would they use crossbar, ccHT link is more likely.

Either way this thing isn't much of a CPU improvement, it just good idea for mobile architecture.

Ho Ho said...

thekhalif
"All of these will add up to 40% per core especially when including L3."

Ok, and besides SSE4, what are unique to Barcelona? From what I know, these listed things are also availiabel on C2D and that will be availiable on the new 45nm CPU's.

thekhalif
"Another major factor I see is that Barcelona is designed for 65nm and switching speed alone may increase by, according to AMD, 40%."

Latest information I've seen said that AMD will have up to 2.6GHz quadcores and up to 2.9GHz dualcores based on the new core. Having faster switching transistors doesn't directly translate to faster CPU's. Also as we are talking about per-core clock-to-clock speed then it is completely irrevelant. If you want to factor in clock speed then be aware that Intel will have 3.7GHz quadcores before this year ends.

Also, adding L3 to a CPU with IMC doesn't help all that much. Compared to Intel they already have half the memory latency and having L3 won't have nearly as good effect as on Intel.

S said...

thekhalif,

Thanks for the pic. But I can't see anywhere in the pic it is AMD chip let alone 65nm Barcelona.

Hv there been any more demos after the one where AMD demoed it on task manager ? Intel has demoed penryn running vista, games etc. I am not taking sides here but, to me it looks like Penryn is closer to release than Barcelona.

Ho Ho said...

Scientia
"How would you know that? The graph doesn't say what the timeframe is. The marks could be beginning, middle, or end of quarter."

Would it matter? If yes then how?

enumae said...

Scientia
How would you know that?

I think it would be a fair assumption that they are trying to show a 3 year time frame, and thus the end of each year.

Here is what the image looked like after I brought it into Autocad, scaled it and traced the graphs.

The numbers are accurate.

PS: Nice job by AMD to gain Mobile and Desktop market share.

Fujiyama said...

Enumae,
if you belive that this graph is correct - AMD is planning to produce and sell 40% more CPUs in 2007 than in 2006.
So AMD is going to end year with 35 points of market share?
Difficult to believe that.

Ho Ho said...

fujiyama
"So AMD is going to end year with 35 points of market share?
Difficult to believe that."

Not after you factor in overall demand growth. IIRC, it was 10% per year.

enumae said...

Fujiyama
if you belive that this graph is correct - AMD is planning to produce and sell 40% more CPUs in 2007 than in 2006...

You are not taking into account that in Q4 2006 AMD was only at 30% Dual Cores.

TheKhalif said...


Ok, and besides SSE4, what are unique to Barcelona? From what I know, these listed things are also availiabel on C2D and that will be availiable on the new 45nm CPU's.


Latest information I've seen said that AMD will have up to 2.6GHz quadcores and up to 2.9GHz dualcores based on the new core. Having faster switching transistors doesn't directly translate to faster CPU's. Also as we are talking about per-core clock-to-clock speed then it is completely irrevelant. If you want to factor in clock speed then be aware that Intel will have 3.7GHz quadcores before this year ends.

Also, adding L3 to a CPU with IMC doesn't help all that much. Compared to Intel they already have half the memory latency and having L3 won't have nearly as good effect as on Intel.


First, I am not comparing to Core2, I am noting the strengths that Barcelona has over K8.

What you are implying is that they can't improve K8 with anything.

Why would they add to the die size if it doesn't help?

Do you think AMD's engineers are amateurs and you know more about increasing IMC perf than they do?

Ho Ho said...

thekalif
"What you are implying is that they can't improve K8 with anything."

No, I'm just saying Barcelona won't be much faster than Core2Duo. In fact, I expect them to be equal when no architectural bottlenecks are hit when benchmarking.

thekalif
"Why would they add to the die size if it doesn't help?"

One thing that might increase die area quite a bit is two IMC's (DDR2/3). I would also think that having wider channels inside and outside of the chip would increase the size quite a bit. The rest of the diespace is probably taken by the other feature improvements. Besides from what I've seen so far the diespace left for the logicks isn't that much bigger compared to K8.

thekalif
"Do you think AMD's engineers are amateurs and you know more about increasing IMC perf than they do?"

Of cource not. My best guess is that dualchannel DDR2 won't be fast enough for Barcelona and they need to lower the load so they wouldn't hit bandwidth limits. Of cource with DDR3 things should be a bit better.

Scientia from AMDZone said...

ho ho
No, I'm just saying Barcelona won't be much faster than Core2Duo. In fact, I expect them to be equal when no architectural bottlenecks are hit when benchmarking.


I would agree. I think Barcelona could still be slower at the same clock, maybe 5%. However, it could do better than I expect and match C2D's speed.

One thing that might increase die area quite a bit is two IMC's (DDR2/3).

What are you talking about? There are not two IMC's. The current IMC runs two channels synchronously and in Barcelona the two channels are split so that they can access memory separately. The change in the die size of the IMC would be insignificant. The changes in the memory scheduler are probably larger.

I would also think that having wider channels inside and outside of the chip would increase the size quite a bit.

No. Processors are made in layers just like motherboards. The change in die size would be very small.

The rest of the diespace is probably taken by the other feature improvements. Besides from what I've seen so far the diespace left for the logicks isn't that much bigger compared to K8.

Excuse me, but what are "logics"?

Of cource not. My best guess is that dualchannel DDR2 won't be fast enough for Barcelona

No, there should be enough.

and they need to lower the load so they wouldn't hit bandwidth limits.

No, this isn't likely at all. The initial bandwidth is just a function of a new design and new transistors; there is no artificial lowering of speed.

Of cource with DDR3 things should be a bit better.

Really? Is this why AMD is petitioning JEDEC to extend DDR2?

Ho Ho said...

Scientia
"What are you talking about?"

Didn't Barcelona support DDR2 and 3? A few months ago it was said it'll have two IMC's, one for each. If this is not so then I'm mistaken (a bit more of this later).

Scientia
"No. Processors are made in layers just like motherboards. The change in die size would be very small."

IIRC, they doubled L1 to L2 bandwidth by widening the channel between them. Did they just added the neccesary lanes to the another layer? To me it doesn't sound logical. Let's say they have 64 lanes between L1 and L2. These are on 8 layers, 8 on each. In new CPU they will have 128 lanes. Unless they intend to add eight more layers I can't see any other way to fit them somewhere but to put twice as much of them to each layer thus doubling the die space taken by them.

Among the other channels I was talking about are double-width HT links, logically they should take a bit more die space. Perhaps not much but more anyways.

Scientia
"Excuse me, but what are "logics"?"

In this context, everything that isn't cache. And please excuse the typo I did before, English isn't my native language :)

Scientia
"No, there should be enough."

Enough of bandwidth? DDR2 800 for four cores means the bandwidth of singlechannel DDR1 400 per core. Not all that much I'd say. Intel has around the same bandwidth as single channel DDR1 333 per core when it uses 1333MHz FSB and enough memory to feed it. Of cource AMD will have a bit higher efficiency thanks to the MC being integrated. With higher memory speeds this will of cource not be as big problem.

Scientia
"No, this isn't likely at all. The initial bandwidth is just a function of a new design and new transistors; there is no artificial lowering of speed."

Who was talking about lowering the speed? I was trying to say that adding more cache will make cache hits more frequent meaning less memory access meaning less memory bandwidth used.

Scientia
"Really? Is this why AMD is petitioning JEDEC to extend DDR2?"

Perhaps because producers are creating modules running at 1333MHz whereas officially fastest DDR2 should be 800MHz. Then again I don't think many people would like to have this kind of modules in their PC: http://images.thgweb.de/2006/03/13/cebit_2006_grafikkarten_motherboards_ram_kuehler_netzteile/geil_memory_3.jpg
http://www.thermaltake.com/product/Cooler/VGAnCHIPnMEM/Memory_cooler/cl-r0026/cl-r0026.asp
http://www.teamgroup.com.tw/xtreem/overclocking/team-xtreem-ddr2-1200mhz-cl-5%115%115%1115/

Also DDR2 should officially run at 1.8V whereas such high-clocked modules often run at 2.3V or more.



From what I know, Barcelona will have DDR3 support when it moves to socket AM3. Does it have some kind of a hybrid memory controller that supports both or will there later be new CPU's that support only DDR3 and the first ones will only support DDR2? As I said, my confusion comes from the information I read somewhere a few months ago that it'll have two IMC's, one for DDR2 and one for DDR3. Also IIRC, it will have IMC per core whereas current dualcores should have one per die.

Rabbit said...

"Hv there been any more demos after the one where AMD demoed it on task manager ? Intel has demoed penryn running vista, games etc. I am not taking sides here but, to me it looks like Penryn is closer to release than Barcelona."

I hope not, Penryn has been pushed back to Q1 2008

http://www.theinquirer.net/default.aspx?article=37334

http://www.digitimes.com/systems/a20070131PD210.html

Scientia from AMDZone said...

ho ho
Didn't Barcelona support DDR2 and 3? A few months ago it was said it'll have two IMC's, one for each.


No. You are getting two different things confused. In 2008, the processors will support both DDR2 and DDR3. Barcelona splits the memory controller's two channels so that they can be accessed separately. There is just one IMC.

IIRC, they doubled L1 to L2 bandwidth by widening the channel between them.

Yes, but this doesn't add that much to the die area. The buses are put in separate layers if room is a problem. K8 uses 9 layers but AMD will use 10 for 45nm. Intel has more manpower to carefully route their buses and connections so they use fewer layers.

Among the other channels I was talking about are double-width HT links

Again, you are getting two different things confused. The HT links are not double width. However, HT 3.0 is capable of splitting a regular 16 bit HT link into two separate 8 bit links. This effectively doubles the number of HT links without increasing the number of HT controllers. The bandwidth is cut in half but that is not much of a factor since the bandwidth has increased by about 4X since HT 1.0.

In this context, everything that isn't cache.

Well, then there is a lot more die space devoted to logic. The SSE hardware is doubled in size, new stack hardware was added, new out of order logic was added, memory access scheduling was made more complex, and more instructions were made fast path. And, of course, new instructions were added plus additional CAS and virualization functions.

And please excuse the typo I did before, English isn't my native language :)

Typos are common; I could tell that you meant logics but just wasn't sure if you meant execution hardware only or were also including microcode.

Enough of bandwidth? DDR2 800 for four cores means the bandwidth of singlechannel DDR1 400 per core.

That should be enough up to at least 2.8Ghz. DDR2 1033 should be good for higher speeds.

From what I know, Barcelona will have DDR3 support when it moves to socket AM3. Does it have some kind of a hybrid memory controller that supports both

Yes. It will have.

Also IIRC, it will have IMC per core whereas current dualcores should have one per die.

No. Just one IMC but the two channels will be split. This does not assign one channel per core. It just gives the scheduler more flexibility.

Scientia from AMDZone said...

There is no comparison between Penryn and Barcelona. The batch of Barcelona in production now should be ready for release.

Penryn is still in the alpha stage and probably includes several bugs. It should take at least two more cycles before Penryn is ready for release.

Remember, first silicon for a new design is usually 1 year before release. Just because a design is in silicon does not mean that it is ready for release. Barcelona is not first silicon; that was in Q2 last year. Penryn is probably the second alpha design and therefore stable enough to demo.

enumae said...

I know we do not agree on this subject, but it would seem that Intel is gonna try and do what I had thought, decrease Dual cores and increase Quad cores.

Here is a link.

Still haven't gotten a response from my last post about the graph.

Any thoughts on what I said about them trying to show 3 year time frame?

pointer said...

Penryn is probably the second alpha design and therefore stable enough to demo.

you didn't read those links? it is A0. Just because AMD only able showed task manager with their not A0 CPU, doesn't means Penryn A0 won't work on office, and adobe.

Scientia from AMDZone said...

enumae

Yes, it does look like Intel will really be pushing quad core on servers. I'm guessing this is because they won't have a new architecture in 2007. This also suggests that they do not have the clock headroom that everyone suggests. If they did, then increasing clock would be much cheaper than doubling cores.

Any thoughts on what I said about them trying to show 3 year time frame?

Well, yes, the graph obviously covers 3 years or 12 quarters. I'm just not certain if the marks are the beginning, middle, or end of each quarter.

Scientia from AMDZone said...

pointer
you didn't read those links? it is A0. Just because AMD only able showed task manager with their not A0 CPU, doesn't means Penryn A0 won't work on office, and adobe.


I'm not quite sure what your point is beyond saying that Penyrn is first rather than second alpha. However, if you truly believe that Penryn is going to be released ahead of Barcelona I'm afraid you'll be very disappointed.

gdp77 said...

Yes, it does look like Intel will really be pushing quad core on servers. I'm guessing this is because they won't have a new architecture in 2007. This also suggests that they do not have the clock headroom that everyone suggests. If they did, then increasing clock would be much cheaper than doubling cores.


Scientia, your observation is right, however your conclusion is wrong... I just bought for my brother's pc:
A) Asus P5B
B) C2D 6600
C) 2x512 DDR2 667 (noname)
for 500€

6600 defaults at 9x266=2.4 GHz

I overclocked out of the box with default voltage at 9x333=3.0 GHz...

Right now the pc is next to me running prime95 for 48hours...

Are u still insisting that they do not have the clock headroom that everyone suggests? Have u actually ever seen a C2D system perform?

Imo, they probably want to "push" quadcores because still doesn't have something similar.

gdp77 said...

Correction :


Imo, they probably want to "push" quadcores because AMD still doesn't have something similar.

enumae said...

Scientia
...I'm guessing this is because they won't have a new architecture in 2007.

I agree.

This also suggests that they do not have the clock headroom that everyone suggests. If they did, then increasing clock would be much cheaper than doubling cores.

Did you think that they might transition to Quad core and raise clock speeds?

Cloverton TDP is 80W for the 2.66GHz, would seem to be some head room if they need it.

Arn't AMD's Quad cores coming out about 2.6GHz at 120W (supposedly)?

pointer said...

repost for missing words :)

Scientia from AMDZone: I'm not quite sure what your point is beyond saying that Penyrn is first rather than second alpha. However, if you truly believe that Penryn is going to be released ahead of Barcelona I'm afraid you'll be very disappointed.


good spinning. I though my english is simply enough for you to understand. why noy you just admit it your bias caused you not believing Intel can have a nice working A0, while AMD cannot?

Did i ever said Penry will be released ahead of Barcelona? It is a well known fact it will not. Nice try to show an unrelated thing as if you were correct on the other statement.

Btw, the release of a chip, sometime is not about its readiness, it is also about marketing decision. AMD can also pull in the Barcelona even the chip is not fully ready (means functional, but lower yield due to process or design).

Scientia from AMDZone said...

gdp77
6600 defaults at 9x266=2.4 GHz

I overclocked out of the box with default voltage at 9x333=3.0 GHz...


I wasn't referring to overclocking; I was talking about clock headroom. Intel's headroom is less than its overclocking ability while AMD's is pretty much the same. This has to do with differences between bulk silicon and SOI. I have not doubt that C2D's overclock better than K8's however this does not mean that Intel can release these higher clock speed. We saw the same thing with both P4 and Dothan. People suggested that the good overclocks proved clock headroom but higher speeds were never released.

Scientia from AMDZone said...

pointer
good spinning. why noy you just admit it your bias caused you not believing Intel can have a nice working A0, while AMD cannot?


You have collections of AMD jokes on your blog and you are suggesting that I am biased? There are no Intel jokes here.

Btw, the release of a chip, sometime is not about its readiness, it is also about marketing decision. AMD can also pull in the Barcelona even the chip is not fully ready (means functional, but lower yield due to process or design).

You obviously do not understand how chips are designed and produced. This is not about process or yield but about debugged and fully functioning hardware. Barcelona cannot be released until it is bug free whether the yield is good or not. You might recall that Intel released Conroe too early and had to deal with an error that hadn't been fixed.

pointer said...

You have collections of AMD jokes on your blog and you are suggesting that I am biased? There are no Intel jokes here.

Another spinning. has or has no intel/amd jokes doesn't has anything to do with this.

You obviously do not understand how chips are designed and produced. This is not about process or yield but about debugged and fully functioning hardware. Barcelona cannot be released until it is bug free whether the yield is good or not.

Too bad, I'm from the related industry and you are obvious not. And you obvious have no idea how the design affect test and manufacturing and hence the yield, don't you? A complex chip can never be bug free but all production chip should be without critical bug, etc. A marketing decision can be made to pull in or push out whenever a company has a functional chip, despite its yield. Pull in with no optimal yield normally is to fight competition, push out even when the chip is ready, possibly to maximize the economical returns (of current offering, much higher yield of new chip, process maturing, etc)

You might recall that Intel released Conroe too early and had to deal with an error that hadn't been fixed.

No, i don't. Another conspiracy theory here? nice try.

gdp77 said...

I have not doubt that C2D's overclock better than K8's however this does not mean that Intel can release these higher clock speed. We saw the same thing with both P4 and Dothan. People suggested that the good overclocks proved clock headroom but higher speeds were never released.

True. P4 could overclock high, but intel never released P4s at these high speeds. Probably this is why Intel knew that P4 was an inferior architecture compared to K8 + P4 was hot. So they didn't even try to release higher clocked P4s.

However this is not the case with C2D architecture. Intel released them at a clock rate that devastates completely K8. That is all Intel needs.

Mark my words: When K8L is released, Intel will respond with C2D @ 3.0 GHz and beyond that.

TheKhalif said...


However this is not the case with C2D architecture. Intel released them at a clock rate that devastates completely K8. That is all Intel needs.

Mark my words: When K8L is released, Intel will respond with C2D @ 3.0 GHz and beyond that.



You Intel guys are amazing. I said Barcelona would be at least 40% faster per core.

I was accused of being a fanboy. All of you can say anything positive about Intel and it's naturally true cause Intel never releases anything slower than an AMD.

And this crap about Task Manager..... what's wrong with demoing a SW load package that shows equal 100% load across all cores.

Task Manager is the only way to see this.

ASSHOLES!!

Scientia from AMDZone said...

pointer
I'm from the related industry


You are a cpu design engineer?

Erlindo said...

AMD's Quads to outperform Intel's by over 40%

Scientia from AMDZone said...

theKhalif
_SSHOLES!!


That's enough of that. This is a warning.

Scientia from AMDZone said...

pointer

Tulsa is proof that Intel still needs P4 and higher clocks on Tulsa would definitely have helped. There were many claims that Dothan would be clocked higher and that Yonah would be clocked higher and neither of these materialized. I'm sorry but overclocking for Intel is not the same as headroom because Intel has greater risk of overheating. I will state right now that we will never see 3.6Ghz on Conroe on 65nm and we'll find out by year's end. What is your prediction?

pointer said...

And this crap about Task Manager..... what's wrong with demoing a SW load package that shows equal 100% load across all cores.


while it is a feat itself to boot to Windows, the task manager doesn't add anything to it (I'll explain later). I'm just using it to compare to what Intel has done with Penryn, A0 running real workload.

if you have done overclocking, or even read an article about it, you will know that the stability is not guarantee even if you boot to OS. The windows might crash at the moment you run the benchmark apps.

Allow me to show you the code that doesn't do much thing and yet eat your cpu cycles.

edit test.bat
@echo off
:loopstart
set x=1
goto loopstart

or if you prefer assembly:
mov cx, 1
notdoinganythinguseful:
inc cx
loop notdoinganythinguseful

run the test.bat and observe the task manager. you can also try to run multiple instance of it if you have multicore.

While it eat your cpu cycles, it doesn't test most part of the cpu, or the I/O

Ho Ho said...

scientia
"I will state right now that we will never see 3.6Ghz on Conroe on 65nm and we'll find out by year's end. What is your prediction?"

I know people with e6400 and faster CPU's have little problems with running them at 3.5GHz+ on air. I can't see a reason why they shouldn't release a faster chip if the feel the need for it.

Of cource it might be that 45nm comes before the need. That way it is quite possible we won't see high-clocked 65nm CPU's.

Scientia from AMDZone said...

ho ho

I'll say again that overclocking is not the same as headroom. It is a fantasy to believe that because you've seen overclocks that high means that Intel could deliver chips at that clock. This is simply false.

Ho Ho said...

Scientia
"I'll say again that overclocking is not the same as headroom."

I can't remember you saing that. Perhaps it was before I starter reading this blog.

Btw, what does it mean that it isn't the same as headroom? Of cource it is the same. Intel could simply bin those better clocking CPU's to be those higher-clocked ones.

If needed Intel could up its TPD from 75W to 125W and release a few 4GHz CPU's too if there would be the need.

TheKhalif said...

_SSHOLES!!

That's enough of that. This is a warning.


I had to get it out. It's the same crap everyday.

Scientia from AMDZone said...

ho ho
Btw, what does it mean that it isn't the same as headroom? Of cource it is the same. Intel could simply bin those better clocking CPU's to be those higher-clocked ones.


No. It is not the same. Intel processors can experience cascade overheating when running at high clocks. This is why Intel put in thermal throttling. For any given batch of chips the tolerance for a given thermal level varies; this is not the same as clock tolerance. If Intel really did just bump up the clock rating to match what most processors are capable of then some of these would be running in a throttled state much of the time. Intel apparently has enough integrity to not do this.

It may not seem fair to you but unfortunately there is a difference between AMD and Intel overclocking. Whatever you can get out of an AMD chip is what you get; the maximum clock rating follows right on the heels of the OC'ing ability. However, to avoid having some chips with a given rating being thermally throttled Intel has to maintain a margin. Therefore the highest clock rating for Intel is not the same as the OC'ing ability. This is the difference between bulk silicon and SOI. BTW, chips that run in a hot environment like engine processors are always SOI to prevent the cascade overheating that I just talked about.

Ho Ho said...

Scientia
"Intel processors can experience cascade overheating when running at high clocks."

That's what better coolers are for. Or do you suggest my e6300@3.1GHz and e4300@3GHz were also constantly throttling? Getting x6800 working at over 3.5GHz is a breeze even with inbox cooler. I know several people who run their e6600/6700's at >3.8GHz on air and not just for benchmarking.

True, not all CPU's can OC that high but for the 0.2% who buy the extreme CPU's there are enough of these.

Scientia from AMDZone said...

ho ho
That's what better coolers are for.


If it were only that simple.

Or do you suggest my e6300@3.1GHz and e4300@3GHz were also constantly throttling? Getting x6800 working at over 3.5GHz is a breeze even with inbox cooler. I know several people who run their e6600/6700's at >3.8GHz on air and not just for benchmarking.

You will not see 3.5Ghz Conroes on 65nm no matter what AMD releases. As I've tried to explain, OC'ing is not headroom.

True, not all CPU's can OC that high but for the 0.2% who buy the extreme CPU's there are enough of these.

Yes, there are. However, you have to understand that the people who overclock can carefully monitor temperatures. They can carefully tweak the voltage levels and use higher velocity fans if necessary. Now compare this with a typical user who doesn't know anything about cpu temperatures or voltage levels so no tweaking will be done. These machines will run with stock fans and this could be in the Summertime in homes with no air conditioning.

In order to really increase clock, the processors have to be capable of running at that speed with no tweaking under less than ideal conditions. This is absolutely not proven when run by experienced OC'ers who test and tweak under good conditions.

Ho Ho said...

Scientia
"You will not see 3.5Ghz Conroes on 65nm no matter what AMD releases"

I agree but not because Intel can't do it but because there wouldn't be any competitior from AMD for 3.5GHz C2D.

I guess 3.33GHz will be the top for 65nm C2D, AMD will likely have 3-3.3GHz dualcore by Christmas.

Scientia
"They can carefully tweak the voltage levels and use higher velocity fans if necessary.

...

These machines will run with stock fans and this could be in the Summertime in homes with no air conditioning."

Did you miss the part where I was talking about increasing TDP and binning CPU's? Also who said you can't put a better stock cooler in the box with a 3.x GHz CPU?

Aguia said...

Intel doesn’t release faster Core 2 Duo, because that would leave the Core 2 Quad far behind in performance VS the Core 2 Duo in most applications.

Not only that but a hypothetical Core 2 Quad at 3.5Ghz would require a TDP of 160W?!

Also the 14-stage pipeline VS the Athlon 12-stage pipeline should allow the Core 2 Duo to clock 16% higher over the Athlon.
So when the Athlon is at 3.0Ghz the Core 2 Duo should be at 3.5Ghz.

Ho Ho said...

aguia
"Not only that but a hypothetical Core 2 Quad at 3.5Ghz would require a TDP of 160W?!"

Well, 2.66GHz Clovertown has TPD of 105W. I don't think 3.5GHz would be anywhere near that 160W.

Intel has shown 3Ghz quadcores in it's 4x4 opponent called V8, I don't think they had too high TDP there. Also I've seen videos of Alan Wake's techdemo running at overclocked Kentsfield running at around 3.7GHz. I don't know what might have been its TDP though but it is certain that it is possible to run those CPU's at high clock rate.

Aguia said...

Ho ho,
As far as I know Intel TDP:

E6700's thermal rating, or TDP, is 65W, while the X6800's is 75W. Fittingly, the QX6700's TDP is exactly twice that of the E6700 at 130W.
The QX6700 runs at “just” 2.66Ghz.

I think that a dual core 3.5Ghz Core 2 Duo will easily go for 90W/100W TDP.
So a hypothetical Core 2 Quad at 3.5Ghz goes well beyond 160W TDP.

Looking at AMD quad max at 130W, Core 2 Duo quad is a power hungry system that will never be built at high clock speeds; the same reason the P4/PD never hit high clock speeds.
That’s the price to pay for a non native quad core processor.

Ho Ho said...

aguia
"Looking at AMD quad max at 130W"
"That’s the price to pay for a non native quad core processor."

Well, that 130W for 2.5GHz quadcore is not much better than Intel's dual-die quadcore at 2.66Ghz so I can't see any benefit from having native quadcore, unless you plan to let them idle, of cource.

Also Intel hasn't yet released its LV Clovertowns. They should release one at 1.86GHz in Q1 and I bet there will be more before this year ends.

Ho Ho said...

Also isn't it interesting that quadcore Xeons don't always have 2x increased TDP compared to dualcores?

Aguia said...

ho ho
They really don’t have the 2X, having just 13% better I wouldn’t call it a big win.
And I was referring to the theoretical TDP of Intel higher clock speed offerings.

If AMD 130W is just for the 2.5Ghz version and not the entire quad core line then I must say that doesn’t look good. Depends a lot what kind of performance it will achieve.

AMD will also have lower clocked quad core processors with 68W TDP.

Ho Ho said...

aguia
"If AMD 130W is just for the 2.5Ghz version and not the entire quad core line then I must say that doesn’t look good"

But that is how it seems. All SE Opterons have 120W TDP, they have frequencies between 2.4-2.6Ghz and release dates from Q307 to Q208.
http://images.dailytech.com/nimage/3796_large_Opty-list.png

aguia
"AMD will also have lower clocked quad core processors with 68W TDP."

Those have clock frequencies of up to 2GHz. Intel has 80W quadcores with up to 2.33GHz and they aren't low voltage versions. LV versions are coming in a little while but I don't know their voltages.

http://www.intel.com/products/processor/xeon/specs.htm

butlimous said...

Thanks for the nice post!

Free PS3