What can a graphics card with AI actually do?

What this slideshow is

Ten full-screen slides that explain what a graphics card with AI does in an office computer: what a token is, what quantising to 4 bits means, how the model reads and writes and how much work it gets through in a day. It works well to show in a meeting or to whoever has to sign off the purchase.

At the top you pick the card, model, quantisation and job, and every figure on the slides is worked out again. Move between slides with the arrow keys or the buttons on each side. The RTX 3090 Ti with Qwen3.8 27B at 4 bits is measured in our lab: 1,158 tokens per second reading and 39.7 writing. The rest are estimates. How every figure is worked out: Method.