Our keyword-spotting accelerator models estimate energy from per-operation costs: an 8-bit multiply, a 32-bit add, a memory read. The numbers came from memory of a well-known 2014 table by Mark Horowitz, and about ten projects said so honestly: "not re-checked".
This run opened the table. It is a figure in the paper, so the values were read off a render of the page. Of 34 constants in our code that have a counterpart there, the largest gap is 3.7%.
The real difference was a modelling choice. We charged every memory read the cost of an 8 KB array. The figure shows read energy rising roughly as the square root of array size. Re-costing 40 designs with that law, extrapolated below 8 KB where the paper has no data point, cuts modelled energy to between a quarter and four fifths. The ranking of designs barely moves (rank correlation 0.991) and 17 of 18 designs on the accuracy-against-energy front stay on it.
The caveats are in the README: the table describes processor caches, not a small weight memory, and only one project was re-costed. Code, data and figure: research/accel-horowitzcheck. Produced with AI assistance (Claude).