Hardware Counters
NVIDIA DCGM hardware counters provide non-intrusive, zero-overhead measurement but previously lacked rigorous validation as an MFU proxy. SM clock sampling noise is bounded and manageable at 30-second collection intervals. Tensor Pipe Acti…
1 sources - 4 claims
NVIDIA DCGM hardware counters provide non-intrusive, zero-overhead measurement but previously lacked rigorous validation as an MFU proxy. SM clock sampling noise is bounded and manageable at 30-second collection intervals. Tensor Pipe Activity reports the fraction of cycles during which Tensor Core floating-point instructions are executing. OFU is specific to NVIDIA GPUs through DCGM and does not generalise to other accelerators.