A utilization percentage is meaningful only with its FLOP and time conventions

Define whether a multiply-add counts as two FLOPs and specify the operations included in forward and backward execution. A simple parameter-based estimate can be convenient for some dense architectures, but it is not an exact law for attention at every sequence length, sparse experts, embeddings, or every optimizer. State how padding, ignored tokens, and variable sequence lengths enter the count. If the numerator is estimated from tokens, make sure those tokens match the work whose elapsed time is measured rather than mixing useful tokens with an unrelated padded-work formula.

For a global completed step, use global model work and the total eligible device peak over that same wall interval. Do not divide a global numerator by one device's peak, or multiply an already per-device arithmetic estimate by device count twice. The peak must match the arithmetic mode being evaluated, including dtype and the convention for structured sparsity. A marketed peak for an unused mode is not a neutral denominator. Heterogeneous devices require the sum of their stated eligible peaks rather than one device peak times rank count.

Activation recomputation can save memory by evaluating some forward operations again during backward execution. Under a useful-model convention, those repeated operations do not increase the numerator simply because the hardware performed them. They can increase a separately defined executed-work or hardware-FLOPs ratio. A configuration can therefore have a higher executed-arithmetic fraction while delivering fewer useful training tokens per second. Compare memory capacity, completed time, and convergence-relevant work alongside the accounting ratios rather than ranking methods by whichever percentage is largest.

Forward and backward cost should come from the chosen model formula or an explicit operation analysis; assuming backward is exactly twice forward is an approximation for particular workloads. Optimizer and auxiliary work may belong outside a paper's model-FLOPs numerator even though their time belongs in the end-to-end denominator. Declare these choices. Profiler FLOP estimates may cover only selected operation families, so adding available estimates is not automatically a complete training computation count. Validate coverage and avoid treating an uncounted operation as if it consumed no arithmetic or time.

Record the model configuration, batch and sequence distribution, accumulation, precision, recomputation policy, hardware peak source, and measured completion boundary. Recompute the numerator when architecture or work changes. If a calculated percentage exceeds one, inspect operation conventions, eligible peaks, device count, timing, and estimator accuracy; do not interpret it as evidence that hardware exceeded a physical arithmetic limit. Compare actual valid token or sample throughput and quality as well. A changed denominator can improve a displayed ratio without accelerating any training work.

The roofline lab uses a related but different synthetic ceiling. Reset, set both I=4 FLOP/byte and B=100 decimal GB/s, predict 400 GFLOP/s, then Play. The fixed compute roof is 1000 GFLOP/s, so its attainable ceiling is 40% of that roof. Reset and set both I=20 and B=100 to reach the 1000 cap. These ratios are ceilings for selected hypothetical intensities, not measured MFU. The instrument has no model work, forward/backward split, device count, tokens, recomputation, or completed-step timing from which real training utilization could be calculated.