Serverless GPU Benchmarks

How to read these numbers

Each row is the median over a campaign's successful runs of a pinned workload — currently parler-tts-mini, a short text-to-speech generation — under a declared scenario, measured from a fixed origin. Real-time factor is inference seconds per second of generated audio — below 1.0× is faster than real time.

• A cold observed scenario includes the observed cold start, evidenced by container session data rather than latency guesses.
• Median startup is the workload's own measurement of application initialization inside the container — loading model weights and starting the app. Provider-side scheduling and container provisioning are not yet attributed separately; they are part of the median total.
• Compare rows only when the workload version, scenario, and measurement origin match.
• A missing cost is unknown, not free.

Full methodology, filters, and API access →