DL boost / AVX-512 / Tensorflow / Windows

[JStewart]

Is it correct that the AVX-512 instruction set is not possible on Windows at this time? It works fine in Linux for me in a VM.

Or is it just a matter of building tensorflow from source? 

I'm unable to find anything on this in the documentation other than that oneAPI isn't available on Windows (which appears to be out-of-date). 

I am getting AVX and AVX2, however. I'm currently on SkyLake-X (which does not have VNNI), but I'm evaluating my choice of CPU upgrade at this time. As I primarily wish to work in Windows I don't see the value in purchasing a processor with that tech if it's not going to be supported.

Thank you.

[AbhijeetJ_Intel]

Hi,

Thank you for posting in Intel Communities.

We are working on this internally and will get back to you soon.

Regards

Abhijeet

[AbhijeetJ_Intel]

Hi,

Normal TensorFlow build support avx512 via oneDNN.

oneDNN uses just-in-time compilation (JIT) to generate optimal code according to your environment.

You can install TensorFlow with pip.

After TF 2.5, using stock TF could also have oneDNN optimization via an environment variable

TF_ENABLE_ONEDNN_OPTS=1

Useful links:

https://github.com/oneapi-src/oneAPI-samples/blob/master/Libraries/oneDNN/tutorials/tutorial_analyze_isa_with_dispatcher_control.ipynb

https://oneapi-src.github.io/oneDNN/dev_guide_inspecting_jit.html

Regards

[JStewart]

I'm sorry, but, given that only AVX2 is reported as supported and the response you just gave me here -- which is applicable to a LINUX ENVIRONMENT and NOT WINDOWS....

..... I'm left to conclude that when you say it generates optimal code for my environment that you are technically correct. It does. And you do not support the DL Boost instruction set or AVX-512 at all on Windows. 

Meaning I either develop in Linux or I waste money on an unsupported processor.

[Ying_H_Intel]

Hi

Just clarify, Intel optitmized tensorflow do support the DL Boost instruction set or AVX-512 at all on Windows. 

here is some doc regarding this:

Intel® Optimization for TensorFlow* Installation Guide

Note: If your machine has AVX-512 instruction set supported, please download and install the wheel file with AVX-512 as minimum required instruction set from the table above, otherwise download and install the wheel without AVX-512. All Intel TensorFlow binaries are optimized with oneAPI Deep Neural Network Library (oneDNN), which will use the AVX2 or AVX512F FMA etc CPU instructions automatically in performance-critical operations based on the supported Instruction sets on your machine for both Windows and Linux OS. 

Thanks

Ying H.

oneAPI AI Support

[JStewart]

I don't understand support here. You and the documentation live in entirely separate universes. Mind you, I cannot even install the AVX512 wheel as instructed because it tells me my platform is not supported. 

And that page literally says the AVX512 wheel is for LINUX ONLY. 

[OOOUPS]

JStewart

You do not need this wheel on windows. 
Support of AVX-512 on Windows is already in main wheel.