Laboratory
Cross any model with any machine
A single grid with every card and every model. You choose what each cell shows —speed, units a day, cost of electricity, memory needed— and the whole grid recalculates. Click a cell to see the full breakdown.
What comes out of this grid
Four conclusions with the filters as they are right now
The grid
Everything against everything
What is measured and what is estimated
The vast majority of the cells in this grid are ESTIMATED: they come from a physical model (each card's memory bandwidth and compute) calibrated card by card with public benchmarks and corrected against the reference measurement. MEASURED cells carry a dot beside them and their source in the breakdown.
MoE models (mixture of experts) are the most uncertain: only part of their parameters fires on each token, and how much of that each engine exploits varies a lot. Prices and electricity costs are indicative. What is measured here is speed, never the quality of the answer.
Read the full methodHow to read the grid
Each row is a card or a machine, and each column an AI model. At the top you pick the figure shown in the cells: speed, units of work a day, electricity cost or the memory needed. Where it says “does not fit”, the card does not have enough memory to load that model at that quantisation.
It is for a quick comparison. A bigger model does not always do your job better, and what matters is whether the machine keeps up with your daily volume. Every figure is estimated except the RTX 3090 Ti's, which is measured. How every figure is worked out: Method.