Question about percentage of FP32 peak we get on H100 and A100 for miniBUDE

Hello, We were comparing the percentage of FP32 peak we get on H100 and A100 for miniBUDE. With the `big5` input (and similarly for the `bm_long` input) we've been seeing results like: | | theoretical FP32| miniBUDE| percent of FP32 peak| | --- | --- |---|---| |H100 | 67|20| 30%| |A100| 19.5| 12|62%| We were expecting similar percentage of peaks on both A100 and H100 if the input size was big enough to saturate the GPU. Our best guess right now is that `big5` / `bm_long` inputs aren't big enough to saturate the H100 as it's much larger than the A100. Does this match with your understanding? If so, are there any suggestions on a bigger input? Thanks!