Many will have seen it already, but I figured posting a link here would not hurt. Another paper (arXiv preprint) in the long line of GPU microarchitecture reverse-engineering efforts:
On quick perusal I did not spot anything particularly surprising.
Curious they chose to compare a data centre class GPU, (H100), with the consumer class GB203.
I guess they used whatever was available to them without having to incur cost for acquiring additional hardware not already available in house.