Yes, indeed supporting OpenCL seems to be reasonable as it can run almost on all GPUs while CUDA works only on NVIDIA cards but besides that, there are not many advantages of it and 65% of our users are working with NVIDIA cards. But let me give you more insights into this topic.
As for now, OpenCL is not well supported and easy to program and the development on it isn’t that active as on CUDA. NVIDIA has incorporated a lot of changes in CUDA to support their latest architectures, which you will miss with the OpenCL. The reason that OpenCL initiative is inert, might be that the companies are starting to be heavy competitors so they don’t want to agree on any standards. Also Apple deprecated support for OpenCL. So the continuation of OpenCL seems to be a question mark.
CUDA is more developer friendly (has better tooling side and the trust libraries) what matters if there is a threat that using OpenCL we can end up with something not optimized for any of the GPUs. NVIDIA has most of the market share among GPU providers and we have recommended NVIDIA so far because of our CUDA implementation within the software.
About CUDA, CPU ratio and more details on which algorithms have CUDA versions and what performances to expect I would need to check with other teams so right now I can’t give you any reliable statistics.
So there is a question what about other 35% of people who are only able to use the CPU when computing the models as still is a number. What I can say is that our team is aware of this situation and we will consider what we can do about it in our future developments and solutions.
What I can do for you is to speak with them on your behalf so any suggestions and your opinion is very welcome as this discussion is still ongoing