ermig1979 / Simd

C++ image processing and machine learning library with using of SIMD: SSE, AVX, AVX-512, AMX for x86/x64, NEON for ARM.
http://ermig1979.github.io/Simd
MIT License
2.06k stars 412 forks source link

Release benchmarks for common CNN networks. #127

Closed Timen closed 4 years ago

Timen commented 4 years ago

I was wondering there is a list available somewhere showing benchmarks for common networks such as Resnet,mobilenet (v1/2/3), squeezene(x)t etc. ?

The library looks impressive but being able to compare performance for common network backbones helps people choose for this library if it is faster compared to other available inference engines.

ermig1979 commented 4 years ago

Now Simd::Neural is deprecated. I develop other project Synet concerned with inference of trained neural network. Synet has benchmarks to compare it with another frameworks. Simd Library has only low level functions to optimize CNN.

Timen commented 4 years ago

Are the benchmark results of Synet publicly available somewhere? I can't find any references to it on the project page.

ermig1979 commented 4 years ago

There are no results in public domain. I performed private comparison with OpenVINO (single thread for every network inference), Synet is faster up to 30-40%. I can publish some of these results.

Timen commented 4 years ago

That would be great! It would really help comparing Synet/SIMD to other options out there. Even better would be if you include a script to reproduce the benchmark results with the actual results so people can easily reproduce the results and possibly run their own benchmarks on their specific deployment hardware.

ermig1979 commented 4 years ago

There is a script perf.sh in Synet to perform this benchmark.

Timen commented 4 years ago

Ok thanks! I'll have a look but nonetheless it may still be handy to have the benchmark results available publicly so one can look at them without having to build the project :)