YC is wrong about LLMs for chip design

The bottleneck for LLM is fast and large memory, not compute power.

Whoever is recommending investing in better chip(ALU) design hasn't done even a basic analysis of the problem.

Tokens per second = memory bandwidth divided by model size.