A good rule of thumb is that a prefill token is about 1/6th the compute cost of decode token, and that you can get about 15k prefill tokens a second on Llama3 8B on a single H100. Bigger models will require more compute per token, and quantization like FP8 or FP4 will require less.