In the current generation of graphics cards AMD gives you more bang for buck in the <1000$ region, because the current Nvidia cards are optimized for gaming in a way that their GPGPU benchmarks lack behind enormously. You also get beter portability with OpenCL because you can run on CPUs, AMD GPUs, Nvidia GPUs, ARM CPUs and GPUs and soon FPGAs. But as far as I can see for most ML on the GPU you are still stuck with CUDA.
Do you know of any open source CNN implementation for OpenCL or an easily adaptlable codebase. I'm also interested in other ML algorithms with Open Source implementations using OpenCL. Or just give me some general comments on the state of OpenCL vs CUDA in the world of ML.
[link][4 comments]