LLM System Design Interview, issue 58, Sep 2, 2026

The FP8 Quantization Tax

Senior ML Systems Engineer interview at Meta, and the interviewer asks:

You moved your matmuls to FP8 on H100s expecting a 2× speedup. You got 25%. Is FP8 overhyped, or did you do it wrong?

Don’t say: FP8 halves the bytes, so the tensor cores should be twice as fast. Maybe the kernel isn’t optimized.

Why doubling Tensor Core throughput silently gets eaten by memory-bound casts and Amdahl's Law — and how elite teams split E4M3 and E5M2 across the backward pass.

The full answer, with the mechanism and the arithmetic, is for paid subscribers on Substack.

Read it on Substack

Get the next one

Free on Substack. Unsubscribe in one click.