Community
cancel
Showing results for 
Search instead for 
Did you mean: 
Gouveia__César
New Contributor I
1,041 Views

Emulating AVX on older CPU in order to use OpenVINO

Hi,

I have made a previous post on if I could run OpenVINO on a CPU which does not support AVX (post) and the answer was:

If you're installing OpenVINO on a CPU that does not support AVX instructions, the installation will fail. As per the system requirements, a 6th to 10th generation Intel Core, Intel Xeon processor, Pentium® processor N4200/5, N3350/5, N3450/5 with Intel® HD Graphics is required. There aren't any workarounds to this that are verified by Intel. 

However, what if I try to use an AVX emulator like this one provided by Intel? 

https://software.intel.com/en-us/articles/pre-release-license-agreement-for-intel-software-developme...

It will work? This kind of AVX emulator has ever been tested with OpenVINO?

 

Thanks,

César.

0 Kudos
3 Replies
Sahira_Intel
Moderator
1,041 Views

Hi César,

Tensorflow provides prebuilt binaries with AVX instructions enabled (so when installing OpenVINO on a system that does not support AVX instructions, Tensorflow is the limiting factor). Using an AVX emulator has not been tested with OpenVINO, but you might be able to do it. 

Alternatively, you can convert your model to IR on a system that supports AVX instructions, the run inference on one that does not. You can also try compiling Tensorflow binaries from source if you cannot use a CPU that supports AVX. 

I hope this information is helpful.

Sincerely,

Sahira 

 

Gouveia__César
New Contributor I
1,041 Views

Hi Sahira,

Tensorflow provides prebuilt binaries with AVX instructions enabled (so when installing OpenVINO on a system that does not support AVX instructions, Tensorflow is the limiting factor). Using an AVX emulator has not been tested with OpenVINO, but you might be able to do it. 

I'm using MXNet instead of Tensorflow so I don't know if this applies to me. Moreover, I know that you can install a tensorflow version that does not use AVX.

 

Alternatively, you can convert your model to IR on a system that supports AVX instructions, the run inference on one that does not. 

How is this possible if the HAVE_AVX flag is defined at compile time???

 

Thanks,

César.

Gouveia__César
New Contributor I
1,041 Views

Hi Sahira, 

Can you please take a look at my answer?

Thanks in advance, 

César.

Reply