SM90a introduces major changes compared to SM120: warp group MMA (WGMMA), Tensor Memory Accelerator (TMA), warp specialization, and thread block clusters.
These features produce very different SASS from what we document today on SM120. Covering SM90a is high priority because the H100 is the dominant datacenter GPU right now.
What is needed:
Access to an H100 or any SM90a GPU.
Running controlled kernels that isolate WGMMA, TMA, and cluster behavior.
SASS dumps with nvcc version and compilation flags documented.
If you have H100 access even occasionally, comment here and we can design the kernel set together.
SM90a introduces major changes compared to SM120: warp group MMA (WGMMA), Tensor Memory Accelerator (TMA), warp specialization, and thread block clusters.
These features produce very different SASS from what we document today on SM120. Covering SM90a is high priority because the H100 is the dominant datacenter GPU right now.
What is needed:
Access to an H100 or any SM90a GPU.
Running controlled kernels that isolate WGMMA, TMA, and cluster behavior.
SASS dumps with nvcc version and compilation flags documented.
If you have H100 access even occasionally, comment here and we can design the kernel set together.