3 ms·
I don't think it has anything to do with the driver itself. At least in my experience, the custom code you write performs more or less the same across both CUDA
by pavanky 10y ago
I don't think it has anything to do with the driver itself. At least in my experience, the custom code you write performs more or less the same across both CUDA and OpenCL. The issues arise when you are using the toolkit libraries (like BLAS and FFT). cuBLAS and cuFFT are much much tuned for NVIDIA hardware than any OpenCL versions available at the moment.
- llukas 10y agoAre there any pain points for cuBLAS and cuFFT?
- pavanky 10y agoNot really, but they are only supported in CUDA. For OpenCL you will have to use clBLAs and clFFT (written and maintained by AMD). This means they are not exactly tuned for other architectures.
- dragandj 10y agoI prefer to use OpenCL, but I think the second key factor why Nvidia dominates (the first is they integrated tooling) is that cuBLAS, cuFFT, cuDNN et.al work out of the box when you install CUDA Toolkit. clBLAS and clFFT? Good luck using them even on AMD hardware, yet alone on someone else's. And, of course, there is no cuDNN equivalent (that is beyond experimental)...
- AlphaSite 10y agoWhat about rocBLAS and rocFFT?