Advanced Micro Devices Inc (AMD)

[wbmw]

Dacaw, you found an interesting link, but did you draw any important conclusions?

Re: "The SPEC 2000 benchmarks are subject to much debate in the scientific community. Are they broken? Do they just depend on memory bandwidth? Do they fit entirely in the cache?"

It looks like the answer is "no".

Looking at SPECint, it looks like the largest gains came from 181.mcf, which has a memory footprint of 190MB. 300.twolf is interesting, since it has a 3.4M memory footprint, which will fit in a 1.5GHz/6M Madison cache, but not a 1.0GHz/3M McKinley cache. Even so, fitting entirely in cache yields a 46% improvement, while the average is 62%. Obviously, fitting into cache didn't help at all.

In SPECfp, the largest gains came from 178.galgel, which has a 63MB memory footprint. Meanwhile, 179.art has a 3.7MB memory footprint, enough to fit in the Madison cache and not the McKinley cache, yet the improvement is 59%, not that much more than the average of 48%. In fact, it doesn't seem like doubling the cache broke any of the SPECfp benchmarks. All 14 had improvements between 25-67%, which is a very expected distribution.

The only program which does look like it "broke" is 181.mcf, a SPECint application which had a 2.74x gain. Even so, it didn't affect the geometric mean all that much. All the other tests ranged between 46-68%.

This article should be a great proof point the next time someone decides to argue the "brokenness" of SPEC.

Re: A four processor 2200 MHz Opteron may reach a similar SPEC2000_rate performance as a four way 1500 MHz Itanium 2 even though the latter has a much higher single processor score

In practice, a 4P Opteron 250 (2.4GHz) does manage to outperform the 1.5GHz Itanium 2 that Hans tested in SPECint_rate (67.7 vs 60.0). The new Madison 9M, however, seems to raise the bar significantly to 72.5. This is likely to be competitive until AMD ramps 90nm. As for SPECfp_rate, AMD's best score of 71.4 still falls short of the 77.9 from HP and 104 from SGI. However, it's clear that with superior memory bandwidth, Opteron can recover much of the gap they started with in the single threaded test.

By the way, I think this means there are a lot of opportunities for Itanium performance to improve as memory bandwidths increase in the future.