Comment by stymaar

For token generation, yes: because current-gen LLMs are autoregressive you need to add the inter-node latency for every since token.

For prompt processing it would work though, and it could for diffusion LLMs as well.