we observed a problem that some convolution layers produce wrong results when executed using a kernel tuning file generated on another GPU.
What we did:
1) Generated a GPU kernels tuning file for our networks on a NUC8 (i3-8109U CPU)
2) Executed the networks on a NUC11 (i3-1115G4 CPU)
What we observed:
Most of the intermediate results are correct, but some of the convolution layers produce wrong results (but consistently the same wrong result). On the NUC8 or when not using the tuning file on NUC11, the results are correct.
Is this issue known?
Thanks for reaching out to us.
There is a known issue about some performance degradations are possible in the GPU plugin on GT3e/GT4e/ICL NUC platforms. Unfortunately, this issue is not addressed yet and there is no workaround for this issue. You may refer to the Release Notes under Known Issues.
Thank you for your question. If you need any additional information from Intel, please submit a new question as this thread is no longer being monitored.