Air cooled Presler XE benchmarked at 4.27 GHz

SINCE GETTING my hands on the 3.46 GHz Presler XE chip late last month, I was somewhat surprised at the relative ease this MCM (multi-chip module)-like two-chip CPU device was overclocking.

That was the case even with the mediocre stock heat sink & fan unit that comes with the D975XBX mainboard. So, before going into the wet world of water coolers, I decided to max it out on the thin air alone, just with a better fan - the wacky Zalman 9500 LED monster with its vertical fan surrounded by a dense ring of sharp, thin copper fins.

The goal in this case was to see how far we could go while keeping stable and reliable operation in benchmarks and stress test, without excessive voltage ups - both while keeping the 1066 FSB but changing the multiplier, and while attempting to speed up the FSB but with a fixed multiplier. After all, two hungry cores on a single FSB do pose some strain on it, so any extra bandwidth to share would be helpful to achieve more balanced system performance.

To play safe and ensure it's not a fluke, i.e, one "special" chip in a thousand, I got the following two setups, each with its own Presler XE; one based on the Intel D975XBX mainboard and Zalman 460W PS; and the other one using Asus P5WD2-E 975X-based board and MGE XG Magnum 500W PS - running both the stock cooler and Zalman one on each board setup, by the way. Both board sets used Corsair TwinX5400UL low latency DDR2 memory running at CL 3-2-2-6 at 533 MHz, CL 3-2-2-7 at 600 MHz or CL 4-3-3-9 at 800 MHz (see photos).

The graphics portion was taken care of by two X1800XT cards, one HIS reference design, another ASUS EAX1800XT Top, a black monster using active cooling with an external laptop-type power supply. For performance testing in 3DMark05 CPU, both were running at 650 MHz GPU/1500 MHz RAM.

The Intel 975XBX platform
By rights, let's start first with the stock Intel cooler that comes with the CPU and mainboard: it seems to be a good heat sink, with large copper surface for CPU. We booted fine at 3.467 GHz - went into Windows, all benchmarks done, but the temperatures were touching 78 deg C at times during stress test and benchmark runs - again, good for extra-cold European winters, but not once the summer comes (or for me in Singapore, the city of eternal summer). Yet there was no throttling on ThrottleWatch.

OK, fine. I upped the multiplier by a notch (14X) to 3.733 GHz - again back to the same Windows benchmarks, the heat sligthly jumped to 80 deg C, but otherwise, there was no throttling. Another jump to 4.0 GHz at default 1.3 v voltage? Booted into Windows, and the heat actually went down to like 77 deg C! Why? Because the CPU started heavily throttling according to ThrottleWatch graphs. In fact, the Sandra2005 results I got in this case were slightly slower than on the standard 3.46 GHz setup. The 4.267 GHz 16X multiplier config also booted Windows, but it did not manage to complete any of the benchmarks, in fact Windows stopped responding after about 3 minutes or so.

Surprisingly, on the same Intel board with Zalman fan, overclocking Presler was a breeze - as long as you enable the "Enhanced Power Slope" in BIOS chipset settings (which I did for the above stock cooler test too). After one hour of trials with different settings, I managed to get the CPU working with 16X multiplier at 4.267 GHz - stable in all benchmarks and stress tests for days without any throttling observed by either ThrottleWatch or speed results - at 1.2875 volts, a voltage slighty BELOW the default 1.3 v out-of-the-box 3.46 GHz setting! Also, the memory worked fine at CL 3-2-2-6 DDR2-533 at 1.8 volts minimum voltage. This means I've got a comparatively cool and "power saving" system at an outrageous speed, at least when measured in raw GHz.

Now, since the 16X setting worked so great without a hitch, why not try the 17X setting for 4.533 GHz clock? No luck, mate - when you select the 17X setting on this board, it instead runs as 1X clock setting, i.e. running the CPU at 266 MHz! OK, that's not exactly a runner, so I reverted back to the 16X multiplier.

Next round was "burn in" (Intel's polite term for overclocking) the system by upping the overall base clock rate for CPU and memory by some per cent - I tried 3% first. Well, it complied, but to no effect - both BIOS and Windows showed the same 4.267 GHz CPU with 1,066 MHz FSB, even if trying to up the voltage by a notch. Next step? Reduce the multiplier to 12X, but force the FSB to 1,333 MHz, another option in the BIOS. Oh yes, it complied too - only for an immediate system hang up upon restart (which was solved gracefully by switching to the maintenance mode). Same results even after upping the FSB and North Bridge voltage - so too bad, this FSB just can't run at 1,333 MHz with three loads (two CPU cores and north bridge).

Back to the "standard" 4.267 GHz setting. With HT enabled, the CPU did produce some record-breaking benchmarks (see table), and, while running four or more heavy tasks at one time (i.e. 3-task stress test plus Sandra 2005 or PCMark05 at the same time, with hardware monitor and ThrottleWatch all on), the temperature never exceeded 75 C as measured. Over 10 days or so of continued use, there were no problems or issues, the platform performed consistently without throttling, and the heat sink fins were never warmer to touch than the (a bit feverish?) human body. In summary, this is (probably) a base for a stable, fast everyday-use platform.

The Asus P5WD2-E platform
Asus' new high-end 975X-based board shows many similarities with the Intel offering in terms of features, except its overclocking options are supposed to be even wider - including setting the FSB clock in 1 MHz increments, for instance. I used another 3.46 GHz Presler XE chip on it as a comparison.

With the standard Intel stock cooler, the board also booted fine at 3.46 and 3.73 GHz in Windows, but the Windows didn't boot at all at 4.0 GHz in this case - however, with Zalman fan, it booted all the way to 4.26 GHz just like the other CPU on the Intel board. That's where the difference comes - it had intermittent problems completing the various benchmarks, sometimes PCMark or 3DMark, sometimes Sandra. I ran ThrottleWatch on it, but it couldn't show the throttling graph at all. Even though I ran the CPU either at default 1.3 v or bit higher 1.312 volts (the 1.287 volts didn't work here), things didn't get better till I disabled HyperThreading in BIOS. Then, at 4.267 GHz, the tests completed fine (see table). Then, up the multiplier to 17X for 4.53 GHz? The system refused to boot back even into BIOS.

Then I tried something else - since the Asus P5WD2-E allows fine-grained FSB clock setting, why don't we try doing what I did sometimes ago with single-core Pentium 4 XE, i.e. a 4,200 MHz CPU on a 1,200 MHz FSB? After all, such a setup should give you better real-life performance than a 4,267 / 1,066 one as the two cores get 12% faster memory access. So, I gave the memory an extra 0.1 v to 1.9 v, and upped the figures - i.e. 14X multiplier and 300 MHz FSB clock. The Windows booted! Now, even though HT was off, the "intermittent" benchmark errors continued, but after a few runs I managed to get correct (presumably throttle-free looking at the numbers) benchmark results for both the 4,200 / 1,200 and 4,270 / 1,216 settings (closest CPU clock match to the 4,267 / 1,066 setup). As you can see, even in synthetic benchmarks, there is a benefit of faster FSB here. Too bad it just wasn't stable.

Talking about stability, I wondered if it's the CPU that is of kind of lower grade than the other one in the Intel board, so I swapped them - guess what? The second CPU performed flawlessly in the Intel board, running stable at 4.26 GHz with HT enabled through all the runs and with similar temperatures as the first CPU, so it is the board after all. I suspect that the early Asus board either had some issues with power delivery (that "enhanced power slope" on Intel board without which you can kiss goodbye to running this beast at 4++ GHz stable) and / or enhanced thermal management support.

Now, back to the Asus board - with fast 1,200 FSB, I upped the multiplier to 15X and voltage to 1.32 volts - and got the system running at 4,500 MHz sharp. It went into Windows smoothly, but, again, froze after starting Sandra. So, looks like non-starter anyway (or, maybe, this speed really needs liquid cooling?). I think the Asus board will probably have some of these issues solved in the next BIOS update, so don't ignore it, OK? Who knows, maybe even 4.8 GHz at 1.2 GHz FSB might be possible then with a good water / phase-change cooler - I'll try it soon anyway, I guess.

A good overclocker
Again, for a dual-chip MCM with two hot CPUs and shared three-load FSB, 3.46 GHz Presler XE overclocks incredibly even on thin air, if you give it a device to dissipate heat into that thin air efficiently (like the Zalman heat sink here). The 23% CPU overclocking, from 3.467 to 4.267 GHz should give you anywhere between 15% and 20% actual speedup on most apps - remember these cores have big 2 MB caches each, so a lot of stuff can be accessed from within at full clock speed. Also, by swapping the CPUs, I proved to myself that the chips themselves seem to consistently scale up to this speed, i.e. while it is not guaranteed of course, many Presler XE's should be able to do this kind of clock speed in stable everyday operation with a top-notch heat sink.

If using good liquid cooling for both CPU and North Bridge on a mainboard with good power delivery and thermals, I believe the 4,500 GHz / 1,200 FSB might not be unattainable goal either - that would give you an 18 GFLOP little desktop supercomputer on a side, and probably the last machine with such boastful clock speeds for some time. Remember, the next one, Conroe, will not run at such clocks any time soon, at least not till its 45 nm - made successor appears sometime around end of next year or so. What doesn't mean it will be slow - besides doing 50% more work per clock, Conroe should come out at least at 3 GHz and, possibly (as Charlie mentioned) at 3.33 GHz for XE variety with 1,333 MHz FSB. How about cooling that one to 4 GHz? µ

CPU	Pentium XE
MHz	3467	4267	4267	4267	4270	4270
FSB	1066	1066	1066	1066	1220	1220
memory	DDR2-533	DDR2-533	DDR2-800	DDR2-533	DDR2-610	DDR2-610
HT	yes	yes	yes	no	no	yes
mainboard	Intel	Intel	Intel	Asus	Asus	Asus
Sandra
CPU int	20337	25026	25036	21421	22698	25104
CPU FP	14098	17277	17317	10270	10894	17281
MM int	49921	61462	61475	45624	48413	61551
MM fp	65946	81209	81252	53873	57146	throttle down
memory int	6692	6631	6754	6243	7554	7576
memory FP	6675	6603	6744	6214	7555	7556