On 2019-06-18 03:15, PICCA Frederic-Emmanuel wrote: > Same here... with WXX100 cards. > what about rocm packaging ?
Not easy. Involves a toolchain. Is ROCm promising enough to challenge CUDA? (Although ROCm has already beaten CUDA in terms of license). Both tensorflow and pytorch supports ROCm. I'm interested in a formal benchmark for comparison with cuda. Scattered user comments are not always consistent to each other. The proprietary CUDA compiler's support for gcc and clang is always lagging behind the latest toolchain state in sid. That makes the maintenance work less interesting. At this point ROCm should be better. I hope AMD could help me get rid of nvidia non-free blobs someday in the future.