GPU Lookup — models that fit a GPU

What fits my GPU — called GPU Lookup in older links — runs the calculator in reverse. Instead of picking a model and getting GPUs, you pick a GPU and get every one of the 141 catalog models that runs on it, with its fit band, speed and cost. It is the page for "we already own this card — what can it serve?"

Inputs

The deck holds six cells: GPU (any card in the sizing catalog, grouped by tier, with quick picks), Precision of the weights, KV cache precision, Context, Users at once and Runtime. Jetson modules are not in this list; their unified memory follows different rules and has its own tool, Edge fit & speed.

The workload behind every row is fixed so rows are comparable: a 512-token prompt and a 256-token answer per request, with the batch equal to the user count. Memory always reserves the full context window for every user — that is the number that decides whether a model fits.

Fits on one card

The main table lists models that fit on a single card within 95 % of its memory, largest footprint first, so the biggest model the card can hold is at the top. Columns:

• Fit and Memory — the fit band and the share of VRAM used. Models at an underused fit (under 50 %) are listed too: they fit with room to spare, which is useful when you want a long context or more users on the same card.
• tok/s — output tokens per second, per user when more than one user is set.
• TTFT — time to first token for the 512-token prompt.
• $/1M — cost per million output tokens at the card's market median price (catalog price when there is no market price).
• Max ctx — the model's native context limit.

Filter chips narrow the list by fit band, by architecture (dense or MoE) and by model provider. Open a row — or tap a card on a phone — and Open in calculator carries the model, precision, context, users and runtime to the calculator for the full breakdown.

Needs more than one card

Models that do not fit on one card are tried on several, within what the card's interconnect allows: tensor parallel 2, 4 or 8 on NVLink and xGMI, up to 4 over PCIe, expert parallel for MoE models, and pipeline stages when no tensor-parallel layout fits. Consumer GeForce cards without a bridge get no multi-GPU rows. The second table shows the minimum GPU count, memory per GPU with its fit band, speed and cost, ordered by fewest GPUs. The rules behind the layout are in How the calculator picks TP / PP / EP.

The summary panel

The panel beside the table answers the usual follow-up questions in one glance: the largest model on one card, the largest dense model at an optimal fit, the fastest model of 30B parameters or more, how many models need two or more GPUs, and how many were hidden because their native context is shorter than the one you set. The size-versus-speed chart plots every fitting model by parameters and tok/s — where a MoE model runs much faster than a dense one of similar size, a caption says so. The share sheet copies a link, a plain-text summary or the JSON; the link restores the whole lookup.