AMD Phenom 9900 Overclocked

WE'VE SEEN a bit of Phenom's performance in the first part this past Friday - how about pushing it a bit further?

This round, I tuned the Phenom 9900 - unlocked CPU, keep in mind - test setup as far as possible within an extra day for fully stable operations under both Windows XP 32-bit and Vista 64-bit, and ran it against a fully tuned Intel QX9770 setup in an "AMD Best vs Intel Best" approach.

One may say that it's not fair since Intel's CPU is more expensive, however, on the other hand, the AMD one can't be bought at all for some time to come.

Both AMD and Intel configurations used dual Asus EAH3870 TOP cards in both default and Crossfire modes, and identical Windows and Catalyst driver versions - so, as much of an apples-to-apples comparison as possible.

Besides the usual 32-bit benchmarks, the 64-bit run added CineBench10 and PovRay 3.7.

The Limits

This is where the Intel and AMD configurations differ in 32-bit and 64-bit modes: Intel could do 4.27 GHz (multiplier 10, FSB1708) in both modes, but the voltage needed for the 64-bit mode was 1.425 volts vs 1.40 volts for the 32-bit mode, if all tests were to complete.

When using Asetek Vapochill freeze cooling, the limits are 4.6 GHz FSB1840 at 1.45 volts Vcpu in 32 bit mode, and 4.5 GHz FSB1800 at 1.4625 volt Vcpu in 64-bit mode. However, this test run only uses the results obtainable on the usual high-end air or water cooling.

In the AMD case, the absolute limit on my setup with similar air or water cooling was 2.88 GHz (14 x FSB206) at 1.325 volts CPU in 32-bit mode, and 2.80 GHz (14 x FSB200) at 1.337 volts CPU in 64-bit mode for a complete test run. Even then, though, the PCMark Vantage 64-bit run couldn't be completed.

The push to 1.35 and 1.363 volts didn't give any improvement except more heat, so there was obviously no point trying harder.

I didn't really bother to push the memory too much here beyond reasonable latency tuning, as CPU speed is the obvious focus of the story - Phenom has no problems with memory performance, as in Sandra cases it still has a slight 10% - 20% advantage in synthetic memory subsystem speed runs over the Intel offerings.

However, even after accounting for its L3 cache exclusivity, Phenom still has only one-third cacheable data size compared to the dual-die Yorkfield and that seemingly more than neutralises its on-die memory controller advantage.

The results

The benchmark results - here they are:

3DMark06 default - Phenom 2880 Crossfire

3DMark06 default - Intel 4270 Crossfire

Povray 3.7 - Phenom 2800

CineBench - Phenom 2800 Crossfire

... and the other results (memory, CPU, Everest) in the table summary

As you can see, the differences are even greater when taking the overclocking headroom into account. It is also very interesting to note the Crossfire scaling: on 2.88 GHz AMD, the total score goes up from 11096 to 15026, while on the 4.27 GHz Intel, the Crossfire jump is from 13516 to 21496... so, it scales better even percentage-wise on the X38 compared to AMD's latest and greatest 790FX chipset?

The memory runs still stand out in AMD's favour, but obviously not for long - in just over two quarters, Bloomfield flavour of Nehalem will close that gap, too.

The Hope

Yeah, hope that this lopsided situation has a chance of balancing out by AMD radically improving either the speed yields of its K10, and/or improving the core IPC by the time the 45 nm shrink is ready.

A combination of slightly lower clock-for-clock performance, and a third lower maximum clock headroom under similar conditions, gives us a very uneven match, where for every 1x of feather-weight Phenom performance, Intel's superheavy-weight Yorkfield provides in excess of 1.7x using the current best-vs-best press sample comparisons.

Basically, they either sell the Phenom 9900 for not more than US$ 300 a piece when it comes out in March or April hopefully, or wave a 'magic wand' to somehow get 3.2 GHz or so speed within the 65 nm process and power envelope if keeping the current microarchitecture intact.

Only at that speed will, in my mind right now, Phenom become somewhat competitive at the high end - and even that only in the US$ 500 - 700 per CPU price bracket. Anything above to fight QX9770? Well, to charge a grand for a Phenom in early 2008, that one better be running at say 3.5 GHz or so.

Well, not all things are that somber. There is one short-term benefit for AMD here - after looking at the today's 3Dmark06 comparison, it does seem to me that the perfect CrossFire platform for AMD GPU sales is there: the Intel X38 / X48 platform running Yorkfield quad cores.

In fact, the QX9770 & X48 official - delayed - February shipping dates match nicely with AMD rollout of HD3870X2 dual-GPU cards. Put two of those in each X48 mobo, CrossFire them with just one bridge cable, and you got a quad-core Yorkfield CPU feeding quad AMD GPU, in just two PCI-E x16 v2 slots! The "Intel Inside" Spider, anyone?

Now, that's a good one for AMD - for every such Intel enthusiast box, AMD would sell some US$1000 or so worth of their high-end GPU card gear... not bad at all, ain't it?

Well, could be better for their finances if they fixed the CPUs, since the margins are way better in that department - looks like Intel will continue to rule here for another year, at least. µ