AI Model FLOPs Utilization (MFU) Calculator

Model FLOPs Utilization answers one question: of all the arithmetic your GPUs coulddo per second, how much is your run actually doing useful model math with? It is the standard efficiency yardstick for large-language-model training, introduced in Google's PaLM paper (Chowdhery et al., 2022), which defines MFU as observed throughput relative to the theoretical maximum throughput of a system running at peak FLOPs.

The calculation is four deterministic steps, all shown in the breakdown above:

1. FLOPs per token. A forward pass costs about 2N FLOPs per token and the backward pass about 4N, where N is the parameter count (Kaplan et al., 2020). So training costs 6N per token and forward-only inference costs 2N.
2. Achieved compute. Multiply FLOPs-per-token by your measured tokens/second: achieved = F_token × throughput. This is the real work your run performs each second.
3. Theoretical peak. peak = GPUs × per-GPU peak, where the per-GPU peak is the dense (non-sparse) tensor-core rate from the NVIDIA datasheet for your GPU and precision. The tool pins these values with a verification date.
4. MFU. MFU = achieved ÷ peak, reported as a percentage. Above 100% is impossible and is flagged as an input error.

The tool verifies the same number two ways — once on the whole-fleet aggregate, and once on a per-GPU basis (throughput and peak both divided by the GPU count) — so you can see they agree. It reports MFU, not HFU: MFU excludes activation-recomputation overhead, so it is always at or below HFU and is the fair metric to compare across different implementations. Use dense peaks only, because real dense-transformer training does not use 2:4 structured sparsity; the sparse peak would understate your MFU by roughly half. For Mixture-of-Experts models, enter the active parameters per token rather than the total.

• Chowdhery et al. — PaLM: Scaling Language Modeling with Pathways (2022) — defines MFU
• Kaplan et al. — Scaling Laws for Neural Language Models (2020) — the C ≈ 6ND convention
• NVIDIA — H100 Tensor Core GPU datasheet (dense BF16/FP8/TF32 peaks)
• NVIDIA — A100 Tensor Core GPU datasheet (dense BF16/TF32 peaks)

The FLOPs-per-token convention (Kaplan 2020 · PaLM 2022 (6N / 2N)) and the GPU dense tensor-core peaks were last cross-checked against these sources on 2026-07-08. The two worked examples above (50.5% and 4.85%) are reproduced exactly by the calculator.