What about radical new compute-in-memory paradigms based on memristors or something? We really need more performance for running transformers. Not for general p

What about radical new compute-in-memory paradigms based on memristors or something? We really need more performance for running transformers. Not for general purpose CPUs.