4 ms·
Me too ;) And with ROCm support for deep learning frameworks such as Tensorflow, GPU price competition aids in "democratizing" AI.
by indescions_2017 9y ago
Me too ;) And with ROCm support for deep learning frameworks such as Tensorflow, GPU price competition aids in "democratizing" AI.
- dogma1138 9y agoTensorflow isn’t supported yet. The ROCm libraries are optimized for AMD hardware explicitly this isn’t democratizing anything it just locks you into a different vendor other than NVIDIA. If anything AMD should follow NVIDIA in supporting OpenACC which is the way to go as far as true probability goes.
- AlphaSite 9y agoDoesn’t ROCm compile to amd’s libs for AMD GPUs and cuda lins for Nvidia gpus? It doesn’t matter that the rocm libs are optimised for AMD.
- dogma1138 9y agoNo, ROCm is an amalgamation of a lot of things and it's been very inconsistently represented, you are thinking of HIP which in theory should that but in practice it's a tool to more easily convert CUDA applications to ROCm, since while you can in theory use NVCC as a backend for HIP the design directive for HIP make it nonviable. In all effectiveness ROCm is AMD's version of CUDA it includes it's own run time and driver stack which effectively replaced the OpenCL frontend for AMD GPUs, basically it gives AMD the ability to do with their GPUs what CUDA allowed NVIDIA to do and it's not to be dependent on a camel (OpenCL). This is what ROCm is: https://rocm.github.io/images/ROCm_Stack.png https://rocm.github.io/images/ROCm_Stack.png or at least this is what AMD is pushing it to be, currently a lot of things are still unfinished or non-existent. The representation of it being open is misleading, while more of the source code is available than for CUDA, ROCm isn't vendor agnostic on any tangible level. In 2-3 years once it's somewhat in a production ready state it will lock you into a single vendor just like CUDA does.