op-mfu-calculator

MFU calculation for matmul requires operator type, tensor shapes, single-run time, and hardware peak FLOPs. It divides actual computed FLOPs by the hardware peak to quantify the performance gap.

MFU in performance analysis measures Machine FLOP Utilization, showing how close actual operations are to hardware limits. You need it for performance baseline evaluation, capacity planning, and diagnosing bottlenecks in matrix multiplication workloads.

Yes, you can compute MFU for both standard and batched matrix multiplication scenarios. The calculation explicitly supports derivations for batched matmul and related attention operators using their specific tensor dimensions.

Matrix multiplication performance falls short of hardware peak FLOPs due to bottlenecks quantified by the MFU metric. By inputting tensor shapes and execution time, you can identify the exact utilization gap and diagnose performance limits.

To calculate Machine FLOP Utilization, you need the operator type, tensor shapes, single-run execution time, and hardware peak FLOPs. If any of these dimensions or metrics are missing, the tool will request them from you.