I am experimenting with binary (1-bit) matrix multiplication on Jetson AGX Orin (GA10B, sm_87).
I have successfully executed the following PTX instruction:mma.sync.aligned.m8n8k128.row.col.s32.b1.b1.s32.and.popc to perform INT1 GEMM.
Nsight Compute reports non-zero utilization on the tensor pipeline: sm__pipe_tensor_cycles_active.avg.pct_of_peak_sustained_elapsed = 7.92%
This suggests that binary MMA instructions (b1.b1.and.popc) are executable on GA10B and are mapped to the tensor pipeline.
However, I could not find any public documentation or datasheet that specifies the peak throughput of binary / INT1 MMA on Jetson AGX Orin.
The official datasheet only lists Sparse INT8 performance (170 TOPS): https://developer.download.nvidia.com/assets/embedded/secure/jetson/agx_orin/Jetson-AGX-Orin-Data-Sheet_DS-10662-001_v1.8.pdf
My questions are:
Is the binary MMA (b1.b1.and.popc) throughput on Orin considered an officially supported and characterized feature?
Is there any documented peak throughput, scaling rule, or recommended method to estimate the peak performance of binary MMA on Jetson AGX Orin?
If not officially specified, is it correct to rely on microbenchmarks and tensor-pipeline utilization to estimate an upper bound?
Any clarification would be greatly appreciated.