Are there any potential benefits of using CUDA instead of OpenCL on Nvidia GPUs? Like, better driver support, ability to utilize Nvidia-specific features?
Nvidia A100 GPU which was used to find a new Mersenne prime has specialized dedicated hardware like tensor cores, which on A100 can work not only for FP16 and FP32 but also for FP64. Are there any benefits of utilizing this capabilities?
Comments
Are there any potential benefits of using CUDA instead of OpenCL on Nvidia GPUs? Like, better driver support, ability to utilize Nvidia-specific features?
Nvidia A100 GPU which was used to find a new Mersenne prime has specialized dedicated hardware like tensor cores, which on A100 can work not only for FP16 and FP32 but also for FP64. Are there any benefits of utilizing this capabilities?
Yes I expect there may be some micro-optimizations that are available on CUDA, such as using bits of PTX in places.
And if the GPU provides some sort of matrix-multiplication on FP64, that we're not currently making use of -- clearly that would be a big opportunity.
But somebody needs to implement it, profile, test.. on some HW.