* Posts by Read Fleming

2 publicly visible posts • joined 27 Jul 2007

AMD: We're not dead yet

Read Fleming

Re: Spec Numbers

> I really don't think you understand any of the numbers you're looking at.

I don't think you read my post carefully.

> First off, (I can't believe I'm having to explain this to somebody), clock speeds don't mean anything between brands. Period. I have a 400mhz MIPS R12000 workstation that renders 80% faster than my 466mhz G3 (IBM 750).

I agree. I wasn't comparing clockspeeds between brands.

> That is perfectly, completely normal. Clock speed simply does not translate to performance once you're comparing unrelated chips.

True when the chips are *unrelated*. In this case, the Barcelona core is closely related to the Opteron's K8 -- see an excellent analysis at realworldtech. The Xeons cited are *the same chip* at slightly different clockspeeds. I neglected to mention the Xenon part numbers, but they're all Core 2 series. I will be the first to say that Intel naming and numbering scheme doesn't make this obvious.

> The other thing is that you ~are~ looking at different chips even though the brand name is the same. The article talks about the new 'Barcelona' AMD processor and the 'Core' architecture based Xeon.

Partially agree. Barcelona is new(ish). To be pendantic, the current Xeons are all Core 2 (I think they finally killed the last of the P4s). Intel did release a very low power "Core" (Yonah-based) Xeon for blades a while back. I don't know if they still sell it.

> The chips you're talking about are the older 'P68' architecture Xeon, and the pre-Barcelona Opteron. There is a huge, huge difference.

Yes, there is a huge difference between the 'P68' Xeon and the Opteron. However, I was talking about the current Core 2 microarchitecture Xeon.

> There is no distortion field here, just a lack of context sprinkled with ignorance.

Perhaps there is lack of context, so let me explain it (Reg editors willing):

If you take a SPEC figure like SPECfp_rate2006 and divide by clock speed, the result is proportional to clocks-per instruction (CPI). The figure is fairly constant across a range of clockspeeds for a given microarchitecture. If you don't believe me, go to the SPEC website and do the calculations yourself. There will also be variations due to compilers and motherboards, but the number is fairly constant. I agree completely that the figure can be wildly different for different microarchtectures, even for the same ISA: P4 vs Core 2 being one example.

My point was that AMD is claiming a 100% increase in CPI for Barcelona, and somehow, running a standard benchmark, they also got a large improvement on the *same* Xeons differing only in the speed bin. SPEC is designed to generate reproduceable results (okay, with a bit of slop). When AMD's numbers are so far off for the Xeon, one has to wonder what else is wrong. Hence the Benchmark Distortion Field, and damage to AMD's credibility.

> For future reference, don't ever bother using clock speed as a performance metric unless you're comparing a chip to itself, like when overclocking.

... or when comparing different speed bins of the same part.

Read Fleming

SPEC Numbers

I just checked the latest published SPEC Rate figures for multiprocessors, and found the following 4-processor results:

AMD Opteron 3.0 GHz: Int = 50.3, FP = 48.6 (16.2/GHz)

Intel Xenon 3.0 GHz: Int = 55.4, FP = 42.9 (14.3/GHz)

According to AMD's Mr. Allen, the 2.0 GHz Barcelona gets an FP score of 69.5. That's 34.8/GHz, double the rate of a pair of current 2-core Opterons. Putting all 4 cores on the same die, adding some cache and some other tweaks should improve the rate, but 100% doesn't seem credible.

On the other hand, being in the proximity of a Barcelona caused a 2.33 Xenon to get an FP score of 54 (23.2/GHz). Maybe AMD has perfected its own Benchmark Distortion Field (tm).