Comment by fwipsy

Right, but datacenter GPUs optimized for LLM training/inference would have a bandwidth:compute ratio scaled to that workload.