[denverllc]

AMD released ROCm (its competitor to CUDA) in 2016, nearly 9 nears after NVidia released CUDA. They relied on OpenCL and failed to invest in "GPGPU", and as a result were so far behind NVidia they couldn't keep up. As a result, for about a decade most scientific GPU code was written in CUDA.

Today, AMD support in PyTorch is minimal. Actually getting anything running is very difficult, and random crashes are common. This is in contrast to NVidia, which spends a lot of money to ensure a full compiler stack and compatibility with AI libraries.

Today, the AMD hardware itself is pretty capable and has a good price/performance ratio. However, actually taking advantage of that performance is difficult because of the poor quality of drivers and software.