[lostmsu]

They use OpenBLAS. PyTorch is slower presumably due to its poor threading. Nothing special about Fortran here, any language without a global lock would have the same result.