Reputation: 1144
I am interested in using the tensor cores from NVIDIA RTX GPUs in python to benefit from its speed up in some scientific computations. Numba is a great library that allows programming kernels for cuda, but I have not found how to use the tensor cores. Can it be done with Numba? If not, what should I use?
Upvotes: 0
Views: 1371
Reputation: 72349
.... I have not found how to use the tensor cores. Can it be done with Numba?
No. Numba presently doesn't have half precision support or tensor core intrinsic functions available in device code.
If not, what should I use?
I think you are going to be stuck with writing kernel code in the native C++ dialect and then using something like PyCUDA to run device code compiled from that C++ dialect.
Upvotes: 2