Dear users,
right now, we have only Nvidia GPUs on our cluster. One interesting competitor could be AMD a it is more and more difficult to buy Nvidia cards and they tend to reserve the newer GPUs to DGX server.
AMD GPUs aren’t using CUDA but ROCm. There is the HIP project aimed to cross compile CUDA to ROCm almost automatically.
AMD pre converted some of the major deep learning software to ROCm .
For our future investment in HPC hardware, it would be good to know if there is an interest to buy those kind of GPU cards.
If interested, our HPC vendor can provide a remote access to a server with AMD GPU as a test.
Is there research group already using or thinking to use/migrate to ROCm or OpenCL?