Open dorugeber opened 5 years ago
@dorugeber Thanks for this comment, this is really useful and we can recompute the potential FP performance based on the AVX2/512 base frequencies to evaluate this. I also got the Broadwell base frequencies from: https://www.microway.com/knowledge-center-articles/detailed-specifications-of-the-intel-xeon-e5-2600v4-broadwell-ep-processors/
I will update this issue once we have re-run the analysis.
Great. This shouldn't qualitatively change the conclusions, but might as well have the right numbers (and it shouldn't take much time to find the data for a handful of processors!). The other issue is definitely the more important one.
Intel processors (since Haswell, I think) have lower AVX and AVX512 clock speeds than the published base speed.
E.g. per https://en.wikichip.org/wiki/intel/xeon_silver/4116#Frequencies, the Xeon Silver 4116 has an AVX512 base speed of 1.1GHz (with a max all-core turbo of 1.4GHz), vs the non-AVX base clock speed of 2.1GHz.
However, the article computes the potential SP floating point performance using the non-AVX base clock speed while assuming AVX512 instructions (each core doing 32 FLOPS/cycle).