Frequently asked questions

It is how fast a language model generates text on a given GPU — how many tokens (roughly, word pieces) it produces each second while you are chatting with it. A higher number means replies stream faster.

Expect the real figure to land within about 20–40% of the estimate. The same GPU and model vary that much between inference engines, engine versions and serving configuration — no spec sheet predicts that, which is why every figure on this site is published as a range rather than a single number.

Vendor-published throughput is usually measured serving hundreds of requests at once. Our figure is for one person talking to the model, one request at a time — the number that actually describes running a model locally. The two are not comparable.

That the model's weights, its conversation cache and the runtime's own overhead all fit inside the card's usable memory at a given quantisation. A model can fit tightly, comfortably, or not at all — we show which, along with how much headroom is left.

Quantisation compresses a model's weights to save memory, at a small cost to output quality. A model that does not fit at full precision often fits comfortably at a lower quantisation — we always report the best quality that fits a given card.

Thousands of cards from NVIDIA, AMD, Intel and Apple Silicon, from data-centre accelerators down to consumer cards. Coverage and estimate confidence are best on modern NVIDIA hardware, where the underlying inference engines are most mature.

Yes — browsing GPUs, models and every estimate on the site is free, with no account required.

Register a free vendor account and add your listings — they appear directly on the matching card's page. See the become-a-vendor page for how it works and current pricing.