The utilisation number nobody publishes

When a training cluster is described publicly, the figure given is almost always accelerator count. Occasionally peak floating-point throughput is quoted. Achieved utilisation — the fraction of that peak actually converted into training progress — is essentially never published, and it varies enormously.

The range we observe in published systems papers and in reconstructed deployments runs from the high twenties to the high fifties as a percentage of peak, for runs of comparable scale. That is a factor of two. Two clusters with identical hardware can differ by that much in delivered compute, and the difference is invisible in every public description of either.

This has a direct consequence for anyone estimating training compute from infrastructure. If you multiply accelerator count by peak throughput by elapsed time, you are computing an upper bound that may exceed reality by two-fold or more. Comparisons between differently sized clusters compound the error, because utilisation falls as scale rises.

We adjust for this in our own estimates using measured utilisation figures where they exist and a scale-dependent prior where they do not. The prior is published with our methodology. It is not a good substitute for the actual number.