SSE2 Mode Makes Opterons Slower Than Athlon XP?

Published by

A japanese journalist said that an investigation he's conducted shows that using SSE2 multimedia extensions with the AMD Opteron makes the microprocessor slow to a crawl. Takeo Noguchi at ASCII Magazine, analysed the Opteron 242 as a microprocessor for a single chip machine, and said that a 1.6GHz chip, according to AMD slightly outperforms a Xeon/Pentium 4 running at 2.8GHz. He told The Inquirer (full story) that he measured the execution time to encode DV-format AVI files to MPEG, WMV, Divx and Xvid. He found that for encoding DivX and Xvid, Opteron 242's performance is roughly equivalent to a Pentium 4-2.8GHz, which matches the expectation from SPEC int/fp results. But, he added, using well known japanese MPEG 2 encoding software TMPGenc (http://www.pegasys-inc.com), the Opteron's performance is a staggering 30% lower than a Pentium 4 2.8GHz. He said: "At first I supposed it was because TMPGEnc is thoroughly optimised for the Pentium 4". But he checked the results from SiSoftware Sandra 2003, and found to his surprise that its multimedia integer performance is much lower than the Athlon XP, the architectural cousin of the Opteron. His results are as follows: Sandra Multimedia Int Opteron 242 (1.6GHz) 6300 Pentium 4 2.8GHz 11148 Athlon XP 2600+ 11614 He claimed that even degrading the Athlon XP to 1.6GHz it will still deliver 8933, meaning, he adds that the Opteron's performance is 30% lower on a clock per clock basis. He said that the results for the Athlon XP come from using MMX multimedia instructions, while the Opteron uses the newly supported SSE2 instructions. He said that means that the SSE 128-bit SIMD instructions are 30% slower than the MMX 64-bit SMD int instructions. He concludes that using SSE2 on the Opteron actually drags down its multimedia integer performance.