PyTorch and GPU compatibility: "Other" GPUs

[vsant]

Hello,

I know that similar questions have already been asked here about this topic and “true” NVIDIA GPU’s compatibility with PyTorch is known (https://developer.nvidia.com/cuda-gpus). The point is that it seems (I am not sure about this) that some manufacturers rely on NVIDIA architecture to create their “own” GPUs. See for instance:

• Galax GeForce RTX 4060 8GB DDR6 128 Bits (https://www.galax.com/pt/graphics-card/1clickoc-series/geforce-rtx-4060-oc-2x.html);
• ASUS Dual GeForce RTX™ 4060 OC Edition 8GB GDDR6 (https://www.asus.com/motherboards-components/graphics-cards/dual/dual-rtx4060-o8g/).

Can we expect the same compatibility between PyTorch and these “other” GPUs? If yes, do we have the same compute capability considering these “other” GPUs and the “true” NVIDIA GPUs (in the case of the “true” NVIDIA RTX 4060 the compute capability is 8.9)?

Thank you.

[ptrblck]

The compute capability won’t change, i.e. a 4060 will have a compute capability of 8.9 as it won’t depend on the actual manufacturer.

[vsant]

Thank you @ptrblck. And to be sure, PyTorch is compatible with these other GPUs as well. Right?

[ptrblck]

Yes, all current PyTorch binaries support sm_89.

[vsant]

Thank you again @ptrblck.