MKL FFT performance reduction using new CPU which supports avx-512

[cyfgod]

Hi，

    We use MKL to do FFT for image reconstruction.

    Previously, based on the Intel E5 2670V2 CPU (Base frequency - 2.5G, 10 cores and 20 threads,  supports AVX instruction set), than switch to new Intel Silver 4314 CPU (Base frequency - 2.4G, 16 cores and 32 threads, supports AVX/AVX2/AVX512 instruction set), It was found that some higher-order algorithms were very slow, more than twice as slow.

    Will it be the impact of the AVX512 instruction set?

Thanks

[Mark_L_Intel]

@cyfgod , the answer on your question: yes, it is possible to see slowdown with AVX-512. If the code cannot be easily vectorized the use of AVX-512 might not provide any performance gains.

It’s important to benchmark your specific application on your target hardware to determine the actual impact of AVX512 on performance. Profiling tools and careful analysis of the code can help identify if AVX512 is beneficial or detrimental in your case. You could use Intel Advisor which is available as a part of oneAPI Base Toolkit: https://www.intel.com/content/www/us/en/developer/tools/oneapi/base-toolkit-download.html

Also, please see this presentation Optimize for Intel® AVX-512 with or without Intel® AVX-512 hardware.