The fun thing is, for LLM you don’t need perfectly repeatable result. It won’t speed up training but running the chips could be significantly cheaper with that kind of tech. Veritasium had a video about it a couple of years back, before the ai craze.