# Published serverless GPU benchmarks

Returned 0 of 0 accepted target results.

Time to result is measured by the caller from invocation start, before provider submission, through retrieval and validation. Median overhead is the per-run client total minus the workload's own compute phases (preprocessing, inference, postprocessing): the platform's time cost — startup or snapshot restore, queueing, and result retrieval. Current asynchronous-provider profiles poll once per second; older campaigns may use an earlier profile. Provider and workload phases may overlap; with a restored startup, application init predates the request and can exceed the client total. Accepted samples accumulate across campaigns only for the same immutable target, scenario, canonical measurement origin, and startup mode. With one successful sample, a value is a single observation rather than a meaningful distribution. Null cost is unknown, not free.

| Provider | Class | Workload | Scenario | Startup mode | Optimization | Origin | Observed GPU | Samples | Failure rate | Median startup (app init) | Median overhead | Time to result (median) | Median RTF | Median cost | Published |
| --- | --- | --- | --- | --- | --- | --- | --- | ---: | ---: | ---: | ---: | ---: | ---: | ---: | --- |

JSON: https://cloudgpuprices.com/api/v1/benchmarks?gpu=rtx-4000-turing&sort=total_ms&order=asc&limit=50
Manual: https://cloudgpuprices.com/agents#benchmarks