They're trying pretty hard, but Nvidia has spent thousands of man years on this stuff and built ecosystem and community around it. It's not easy. Plus it's hard for AMD to hire the best software folks.

Cerebras is pretty well suited for large language models like GPT3. Their latest generation product can be clustered easily to train huge models. I wouldn't say they're ending AMD and NVDA though, but in order for huge language models to be democratized, some disruptive technologies have to happen. No one other than whales today can afford to train GPT3.

How many APUs can be connected together via IF? Hopefully they can do 8-16 to challenge DGX.