Card profile
Everything
a card does in a day
Pick a card and a model. Below you have, job by job, how much work it gets through in 24 hours straight: how many invoices, how many emails, how many photos, how long each one takes and what the electricity costs.
The three versions of each model: Q4 (4 bits) · Q8 (8 bits) · not quantised (16 bits). Quantising lowers the precision of the weights so the model takes up less memory and runs faster; that is why the same model fits or does not fit depending on the version.
At what pace
One at a time, or all at once?
—
24 hours straight
Everything it does in a day
—
Day planner
An office day
Tell it how much work you throw at it each day and watch how much of the day the card eats.
How busy the card is
—% of the day
—
Compatibility
Which models fit, and in which quantisation
—
Each cell: whether it fits, the memory it needs (weights + context) and the writing speed. The colour follows how it feels to use: red below 5 t/s, amber up to 15, green from 30. The highlighted version is the best one this card can hold.
And now what?
The figures come from the shared Automata Local AI calculation engine. Every number carries its measured or estimated label: nothing here is a made-up figure.
The calculation engine (tw-engine.js) is not there yet. Reload the page in a moment.
What the profile is for
This is the profile of a single graphics card: job by job, how many units it gets through in 24 hours straight, how long each one takes, how many people that equals and how much electricity it uses. It also shows which models fit and at what quantisation. Change the card, model or quantisation at the top and everything is worked out again.
It is useful when you already have a card in mind, or one at home, and want to know how far it goes before giving it real work. The figures are estimated from its specs, except the RTX 3090 Ti's, which are measured. How every figure is worked out: Method.