Methodology & accuracy
How the numbers are built and where they are wrong.
-
01
Methodology & data sources
How VRAM and throughput are computed and where the GPU, model and benchmark data come from.
4 min read · updated 2026-10-26 -
02
Why our tok/s are different
Why the estimates differ from vendor peak figures: batch, context, prefill and realistic bandwidth.
3 min read · updated 2026-10-26 -
03
What is not modeled
The effects the calculators leave out, for inference and fine-tuning.
3 min read · updated 2026-10-26 -
04
What the model actually knows
The architecture facts the model uses per layer, and the ones it has to approximate.
3 min read · updated 2026-10-26 -
05
Accuracy and error bands
How close the estimates are to measurements for memory and throughput, from the public calibration set.
3 min read · updated 2026-10-26