What this path costs, and what it does.
The card says what will launch. This page says what a launch actually did — one row per rented box, every number from a paid hour.
The cells.
Each row is one launch on one box. A field the capture did not record stays empty rather than being inferred from a neighbouring cell.
| hardware | engine | TP | price | KV pool | decode | TTFT | measured |
|---|---|---|---|---|---|---|---|
rtx-pro-6000-x4 | vllm | 4 | $5.49/h | — | — | — | 2026-10 |
Why performance is not on the card.
A card is per weight path. Performance is per cell — the same weights on four cards is a different machine from the same weights on eight. Putting a number on the card would make one cell speak for every box the model could run on.
So the card carries a performance: link and nothing else. This page is
where the measurements accumulate: each launch against this path appends a row, and the
row names the box, the engine, the tensor-parallel width, and the price that hour
actually cost.
A cell that fails is recorded too. A path whose every launch failed is a different answer from a path nobody tried, and this page keeps them distinguishable.