Uncertainty and calibration

An estimate is a calibrated range, not a benchmark promise. The calibration feed records model, hardware, runtime, , , batch behavior, and measured throughput. It filters outliers, separates batch-one runs, and fits per-stack efficiency with p10, p50, and p90 bands. The middle value is the planning estimate; a measured result is still the final truth for one machine.

Formula

MBU = measured tps × bytes read per token ÷ peak bandwidth
error ratio = measured tps ÷ estimated tps

Worked example: The RTX 4090 8B calibration anchor

For an RTX 4090 and an 8B Q4_K_M model, the physics estimate is 153 tps and the artifact's measured band is 140 to 180 tps. The observed range is about 8.5% below to 17.6% above the estimate, which is why the product shows a band rather than false precision.

low residual = 140 ÷ 153 - 1 ≈ -8.5%
high residual = 180 ÷ 153 - 1 ≈ +17.6%
Try a current model on the calibrated RTX 4090