Some thought on Hopper GPU fine grained sharing

G-Safe persitent kernel+context interrupt can do simple sharing for small kernels to replace mps( MPS is implemented SM masking)，AMD has multiple GPU context per SM. Instrumentation of PTX does bound checking.(Can do SM occupation sampling also?)

Arax

Some API exploits from plasticine, not sure still valid today.

A lot of benchmarks open sourced there