Comparing the performance of different versions of the OpenVINO library

[ksv87]

I have two versions of the OpenVINO library: 2021.3.394 and 2022.3.1. I run inference on the same model with the same hardware, without setting any additional parameters.

When running inference on the GPU, the results are comparable, with approximately the same frames per second (FPS), as expected. However, when running on the CPU, the newer version provides a 2x increase in FPS. Why is this?

I need to use the older version due to dependencies in the older part of our software. However, I would like to try and configure it such that the FPS is more comparable. Is this possible?

[ilya-lavrenov]

It's hard to say exactly what is a reason of such improvements.
OpenVINO merges tens of commit every working day to improve quality and performance, so between these versions we have 1 year and hundreds of commits.

dependencies in the older part of our software

Maybe it's better to resolve this issue? OpenVINO has very small amount of dependencies:

• TBB on C++ level
• numpy on Python level.

[ksv87]

Thank you for answer

Maybe it's better to resolve this issue? OpenVINO has very small amount of dependencies

I mean, the old software segment uses the specified version of the library. This segment is being developed by another team and is freezed, so I can't influence it. However, I would like the software to not depend on multiple versions of openvino (2021.3.394 and newer).