Replies to post #86259 on Advanced Micro Devices Inc (AMD)

[Elmer Phud]

Base isn't before Result, its below Result on any single submission.

Would it be too much trouble to simply look at the webpage where they list all the scores?

http://www.spec.org/cpu2006/results/res2008q4/

The last 2 columns say base and peak. Base comes first and peak comes second. So as an example, Intel's SPECint_rate scores for a single socket i7 are 117/125 base and peak.

Now back to the comparison, that 117/125 Intel score is better than the 2 socket score of Shanghai, which is 113/136, when comparing base and almost matches when comparing peak.

Rather remarkable that a single i7 could match 2 Shanghai's, especially when you look at AMD's choice of compilers. They used PGI and Pathscale which everybody knows unlike Intel compilers, are simply designed to produce good SPEC scores on AMD chips and never used by any serious software developers for mainstream applications. Intel compilers are used by Oracle, MySQL, HP, and a host of other commercial and University level scientific applications by the likes of University of California, CERN, Bell Labs, Lucent, Cornell, Oxford, just to name a few.

I think it quite possible that i7's SPECfp_rate scores will show the same ~2X performance lead when released in Q1. Meanwhile we will sit back and watch to see if AMD can actually deliver meaningful volumes of slower Shanghai's on a new unproven 45nm process.

[Elmer Phud]

Pete -

More indications that i7 2X may more than double Shanghai's performance in FP.

In addition, the Nehalem EP was tested with the Stars Euler3D benchmark, which is a computational fluid dynamics simulation that majors on floating point performance. Here, the Nehalem EP managed to complete five instances in 14.34 seconds, considerably better than the 30.32 seconds scored by a pair of 2.7GHz Shanghai processors.

http://news.softpedia.com/news/Intel-Nehalem-EP-Gets-Early-Benchmark-98425.shtml

You have quite a task ahead of you. When Intel thoroughly trounced AMD in SPECint and SPECint_rate you decided they were no longer important benchmarks, and SPECfp_rate was the only true measure. Now it looks like you're going to have to dump SPECfp_rate and come up with another "true measure" of processor performance. Any candidates?