Last week when Tesla announced that all production cars now come standard with an Nvidia Drive PX 2 system, Elon Musk said that it was a "tight call" with AMD in regards to who had the better deep learning hardware. This is a field I am very interested in and it has been dominated by Nvidia. The four major tool-kits for deep neural network (DNN) machine learning are TensorFlow by Google, Caffe, CNTK by Microsoft, and Torch. Unfortunately, none of these run OpenCL without third party modifications, which are not officially supported. My question is what is AMD going to do about this? Tesla obviously tested something and found it to be close to Nvidia's capabilities. However, there is currently no mainstream option. Is OpenCL not the answer? Are HIP or the HCCompiler the answer? Right now, anybody interested in developing a world class machine learning algorithm has no choice but to do so on Nvidia hardware with CUDA. Serious support of one or more of these tool-kits would be helpful.
