We have an A5000 on a machine without AVX, which our platform has been developed on. The A5000 requires CUDA 11 and above, and as far as we know CUDA 11 is compiled with AVX, so we can’t use it. However we can run CUDA 10.2, but we are getting errors related to kernels [TF, PYTORCH]. And we believe this is might be because of SM86 vs. earlier versions on CUDA 10.2. We tried compiling PyTorch and TF with kernels in SM versions supported by 10.2, but we still get the same errors. Can anyone help?
AFAIK, avx is required by cudnn 8.1 and up, not cuda itself. Cuda 10 won’t work on your A5000. I guess your only chance is cuda 11.1 + cudnn 8.05 and then recompile TF/PyTorch without avx.
Thanks - will try this and let you know.