Inference Hardware Leaderboard

Frontier accelerators, ranked on the runs published for them. Throughput rises as the per request rate falls, so a rank means nothing until a point on that curve is chosen. The board opens at 8 chips, at or above 30 output tokens a second per request, and the controls above the table move it: a model, a shape, a deployment size, a chip, the rate a request has to see, and which columns you want.

We went through every published run for these chips, kept the fastest one that clears the rate you set, and linked the run behind each row. Every figure is the official source’s own, used exactly as it was published. The hardware those runs used was supplied by the vendors being ranked, which the source discloses and so does this page.