4 ms·
A bigger problem than the performance is the very poor software support. None of the top frameworks officially support AMD or openCL. Only broken patchsets are
by etaioinshrdlu 7y ago
A bigger problem than the performance is the very poor software support. None of the top frameworks officially support AMD or openCL. Only broken patchsets are available for both tensorflow or pytorch.
Performance is an issue, but the bigger issue is that your model just won't work.
NVIDIA is really the only game in town right now. Google's TPUs are somewhat of an exception but they only sort of count, because you can't just go buy a TPU for training. Makes TPUs a non-starter for a lot of businesses.
What's interesting is the slow pace of change. 3 years ago we were in the exact same place. I don't see it changing overnight.
- slavik81 7y agoApple abandoned OpenCL years ago. It's now officially deprecated on macOS. The most recent version they implemented is OpenCL 1.2 (released 2011). For reference, the current version is OpenCL 2.2 (released 2017).
- mdaniel 7y agobecause you can't just go buy a TPU for training but that's what this is, right? https://coral.withgoogle.com/products/ https://coral.withgoogle.com/products/
- hollerith 7y agoThat's for low-power devices.
- bitL 7y agoIt's both poor software support and poor performance, e.g. Vega 20 has 15TFlops in theory, in practice one could get around 7.5TFlops. And only use models supported by ROCm, custom CUDA kernels used by many models nowadays might not work after cross-compilation to AMD instructions.