It seems TensorFlow only supports CUDA and not OpenCL.
I saw the tensorflow-cl project, which compiles the CUDA code into OpenCL, but it is still a development version which does not work in all cases.
My question is whether Google, TensorFlow's developer, will ever develop a multi-platform version of its tool (no, I do not mean the CPU only version). Are the features of proprietary CUDA so critical to focus on a single GPU vendor? Are there any plans to develop an OpenCL/Vulkan/SPIR-V version at anytime in the future?
The answer is obviously yes, and Tensorflow started supporting OpenCL SYCL in the last few weeks, still in the master github branch and with few basic kernels. Many other kernels need to be written in the new format and contributions are welcome.