ROCM vs CUDA in September 2023?

Posted by tronathan@reddit | LocalLLaMA | View on Reddit | 5 comments

Now that rocm is getting implemented in more libraries, as of right now, is AMD becoming a viable option over Nvidia for LLM inference and training? What can someone *not* do with amd/rocm that they do regularly with nvidia/Cuda? (Assuming 30-series; I’m less concerned about the advantages that are specific to 40-series cards)