67% less kernel code is the more interesting number here — Hopper's async capabilities have been underutilized largely because the programming model is painful. Curious how it handles cases where compute and memory phases aren't cleanly separable.

This seems like a better version of CUDA, for Hopper GPUs?

[dead]