[QST]About data regarding h100 in FP8 is inaccurate.

[hjq0127]

What is your question?
In instance 54 of Cutlass, I tested the FP8 data for PCIe H100, but this data does not reach the official standard. Is there anything I need to adjust? (I know the official value is a theoretical value, but this value is less than half of that.)

[hjq0127]

root@estar:/home/test/test/cutlass/build/examples/54_hoper_fp8_warp_specialized_gemm#./54_hopper_fp8_warp_spe cialized_gemm --m=8192 --n=4096 --k=8192 Disposition: Passed
Problem Size: 8192x4096x8192x1
Rasterization: Heuristic with a maximum CTA swizzle of 1 Avg runtime: 0.796641 ms GFLOPS:690092