AMD EPYC + Radeon Instinct To Power

[pegasus]

Is there a document somewhere explaining how to get ROCm Tensorflow running on a generic Linux system?

Hats off for having a will for setting it up yourself. But kids these days just find a docker for that somewhere and run it

[coder]

It will be interesting to see ROCm on super-computers and a pleasant change of pace compared to the proprietary CUDA/compute used by today's NVIDIA-based supercomputers.

Well, the Intel/Cray Aurora supercomputer should also be sans-CUDA.

Of course, Intel is now touting their oneAPI. Presumably, they're trying to take a page from CUDA's playbook, though I believe their stated goal is to provide CPU/GPU portability.

[coder]

The good news might be deep learning guys are not using CUDA directly.

Nvidia is pushing TensorRT, hard. They are building up more toolchains, services, and even pre-trained models around it, as well. A whole ecosystem, really.

I think they realized their mistake in supporting open frameworks, and are now working hard to replicate CUDA's success in the deep learning domain.