Performance in Synthetic Benchmarks: SSE2 in Athlon 64 FX
Another innovation introduced in Athlon 64 and Athlon 64 FX is the integrated set of SSE2 instructions. Just like in Pentium 4, there is a single set like that here, so it would be really interesting to evaluate how efficient it is. We have already asked this question during our Opteron 144 tests (see our article called ASUS SK8N + AMD Opteron 144: Uniprocessor Workstation on AMD Opteron and NVIDIA nForce3 Professional for details). As you remember, we used BLAS test from ScienceMark 2.0 test set, which allowed us to conclude that SSE2 instructions set in CPUs based on AMD architecture is not so fast, as we would like it to be. Now let’s try to view this problem from a different viewpoint. This time we will take SiSoft Sandra 2003. this test package includes two benchmarks capable of measuring the performance of the SSE2 instructions unit. They are: the well-known Whetstone C 2.0 test rewritten for the SSE2 instructions and the test based on the measurements of time required for Mandelbrot sets calculations. Here are the results for both of them:
As we see, the SSE2 unit of Athlon 64 FX is not that impressive. The performance is quite low here because of the low processor frequency compared with the working frequencies of the Pentium 4 processor, as well as because of the not very successful SSE2 unit implementation. All in all, we can say that SSE2 support is a formal feature of the new AMD processors. AMD engineers seemed not to care that much about high performance of this unit thinking that it was not worth the effort probably. However, AMD has finally managed to catch up with Intel in terms of supported SIMD extensions (here we do not take into account 3DNow!). this situation will probably last for another 2-3 months until Intel finally announces its Prescott CPU, which will boast a few additional SIMD instructions. They will have to do with the translation of the floating point numbers into integers, complex numbers processing and video decoding.