New System

I waited to get enough spare money and then waited because of rumors of the 3.4 Ghz Phenom II 965. So, I decided to stop waiting and put together an interim system.

I didn't really care for the wattage bump on Phenom II 965 to 140 watts. I figure that AMD will probably release a new one at 125 watts like they did with Phenom so I'll wait for that one before getting a quad core. I did briefly consider Intel but unfortunately the Penryn based quad cores are pretty much obsolete because of the FSB bottleneck. That only left i7 920 but I didn't feel like spending all the extra money for little additional CPU power. Also tipping the scale was that I decided to wait for AMD's 5000 series GPU's to see how good they are. But with an Intel motherboard I would be stuck with some truly second rate graphics. Maybe if i5 were out there would be a genuine option from Intel but right now, it doesn't exist. I picked up:

CPU: Phenom II X3 720 Black Edition (tri-core)

Motherboard: Asus M4785TD-V EVO

Memory: 4 x 2GB DDR3 OCZ Gold AMD edition (capable of DDR3-1600)

Power Supply: Corsair HX 850 W Professional Series (modular)

Harddrives: Western Digital 640 GB, 1 TB, and a Western Digital Elements USB 1.5 TB external

DVD drive: Lite On DVD reader

Monitor: Asus 21" LCD

OS: Vista Home Premium and OpenSuse Linux


I'm using the stock HSF so overclocking is limited primarily by cooling. That is fine since I'm not really interested in any severe overclocking. To test stability I run Prime95 with maximum heating. From every test I've done Prime95 always comes up as the toughest. I've found that running windows is not as stressful as running the stability tests in AMD OverDrive and the AOD tests are not as stressful as Prime95. This makes the common practice among overclockers of "being stable enough to get a SuperPi score" something of a joke. I wouldn't trust a system without running the toughest stressor to see exactly where I stand.

The M4A785TD motherboard has had a few BIOS updates. I installed the latest one. I wasn't really interested in trying to unlock the 4th core as some have done since I knew that this would be a weak core. However, others asked me about it so I did give it a try. I was unable to show a 4th core with any configuration using Asus' Unleashed mode. Also, even though the BIOS has an option for setting the NorthBridge multiplier independently of the HT clock it doesn't work. Maybe in the next BIOS update. So, in order to increase the NB speed I have to bump the base clock just as people would have to if they didn't have a Black Edition with unlocked multiplier. I'm running memory in Auto which defaults to ganged mode but I don't know that that makes any difference in stability. Perhaps it would at 1600 but with only three cores I don't really need more than DDR3-1200. With CPU and NorthBridge voltage on auto, memory set to 1.59 volts and 533 Mhz, I ended up with:

250 Mhz base x 13 = 3.25 Ghz is stable. However, the 2.5 Ghz NB is overkill. With all 4 DIMMs I get 667 Mhz with a latency of 7 and 1T command rate.

238 Mhz base x 14 = 3.332 Ghz is stable. The 2.38 Ghz NB is still higher than it needs to be. With all 4 DIMMs I get 634.7 Mhz with a latency of 7 and 1T command rate. If the multiplier were locked this would be the fastest configuration.

231 Mhz base x 14.5 = 3.35 Ghz is stable. The 2.31 Ghz NB is okay. With all 4 DIMMs I get 618 Mhz with a latency of 7 and 1T command rate.

225 Mhz base x 15 = 3.375 Ghz is stable. The 2.25 Ghz NB is about right. With all 4 DIMMs I get 600 Mhz with a latency of 7 and 1T command rate. This one or the following one seem to be about the best all around configurations.

219 Mhz base x 15.5 = 3.395 Ghz is stable. The 2.19 Ghz NB is not too bad. With all 4 DIMMs I get 584 Mhz with a latency of 7 and 1T command rate.

212 Mhz base x 16 = 3.392 Ghz is stable. The 2.12 Ghz NB is about the lowest I would want to go. With all 4 DIMMs I get 565 Mhz with a latency of 7 and 1T command rate.

206 Mhz base x 16.5 = 3.398 Ghz is stable. The 2.06 Ghz NB is a bit slow. With all 4 DIMMs I get 549 Mhz with a latency of 7 and 1T command rate. If I specify 667 Mhz in the BIOS the memory defaults to a latency of 9 and 2T command rate. I might be able to tweak the settings back to 7 and 1T, but with the slower NB it wouldn't be worthwhile.


Update:

i5 750, i7 860, and i7 870 were not out yet when I ordered my components. However, now that I've seen the reviews it is clear that these processors wouldn't have mattered anyway.

Remember back when AMD released the B3 stepping of Phenom in early 2008? AMD discovered something that they had overlooked. Their processor wasn't playing nicely with the 700 series southbridge. So, AMD released the new 750 southbridge; and, when these Phenoms were used with motherboards having the 750 southbridge, it made a difference. You could easily get 200 - 300 more Mhz on an overclock. With that painful lesson under their belts, AMD upgraded the phase lock loop in the 45nm Phenom II's so that they got the same benefit whether you used the new 750 or the old 700 southbridge.

However, Intel has apparently fallen into the same experience trap with the newest processors. These chips have PCI-e on the die itself. Great for reducing cost but not so great for overclocking. Intel ties the PCI-e clock to the BCLK much as AMD has tied the HyperTransport frequency to its base clock since K8. Even Anandtech admits that getting away from 133 Mhz multiples will cause PCI-e problems. In contrast I had no trouble running the base clock up from 200 Mhz to 250 Mhz; HyperTransport still worked fine. Secondly, as you increase the frequency on Intel's newest processors it destabilizes PCI-e because the drive transistors cannot keep up at stock voltage. The solution would normally be to overvolt but unfortunately this can't be done very well with Intel's stock HSF. In fact, Anandtech used the word "sucks" several times in describing overclocking with the stock HSF. Anandtech claims a top clock of 3.37 Ghz for the i7 870 with stock HSF. This would be the same as what I'm getting however given Anandtech's checkered history with testing I have to assume that they did not try running Prime95 on all four cores with maximum heating as I did. This sounds like it would easily knock their claim down to the same 3.2 Ghz stock that Chile Hardware was able to get.

This is a problem for Intel if you really want more performance. Anandtech also says that you need to disable Turbo if you want the maximum clock without crashing. However, they then say that you need to leave Turbo on to let the system clock down the cores individually to avoid wasting power. And, even though Intel quietly suggested that reviewers use the Thermalright MUX 120 premium cooler to solve Lynnfield's thermal woes, Xbit Labs saw temperatures of 93 C under load at 4.07 Ghz. No thanks. Maybe Intel will fix these hardware issues in the newer 32nm i5's and i7's but for now these chips have issues. Nor have I even mentioned the problems encountered in trying to get Turbo to work under Linux or the way Turbo is unlikely to work in a standard case with a nice video card installed.

To get the current speed on my system I only had to change the base frequency, the multiplier, and bump the DIMM voltage to 1.59. The CPU base and NB voltages and DIMM timing are all on auto. Even with the bump in speed however I'm still running both the CPU and memory under their maximum rated voltage. My processor idles at 900 Mhz at 25 C and runs up to 3.375 Ghz with about 48 C under Prime95 load. If Intel were able to do this then they would have something to brag about.

I don't know what some of the reviewers have been thinking but worse overclocking than Bloomfield with 33% less memory bandwidth hardly sounds like a revolution. PCI-e onboard to reduce cost while also sabotaging overclocking sounds more like a reluctant compromise than a feature. And less functionality with a higher price tag than Phenom II is just not my idea of a good deal. Intel has a few obstacles to overcome (or a sharp price cut) before the new i5's and i7's will be genuine competition for AMD's Phenom II's. However for anyone with money to burn, a premium cooler, and a calm disposition the new Lynnfields could provide endless hours of fun trying for a high SuperPi score. I would stay away from Prime95 though at least until Intel comes up with a solution to the PCI-e problem.