Same model.
Same prompts.
Vastly different results.
Apples-to-apples inference benchmarks: the same model on General Compute and on a GPU cloud baseline. All measurements include network overhead.
How we measure.
View GitHub repo >- 01
Identical requests
The same prompts to gpt-oss-120b on both providers: temperature 0, up to 1,000 output tokens, interleaved in the same time window.
- 02
Two prompt sizes
256-token and 1,000-token prompts, 50 runs each. Warm-up runs are discarded, and a unique nonce defeats prompt caching.
- 03
Three metrics
Output throughput after the first token, time to first token and end-to-end latency, network included.
Output throughput
Up to 12.3×
Higher is better
256-token prompts
1,000-token prompts
Time to first token
Up to 7.1×
Lower is better
256-token prompts
1,000-token prompts
End-to-end latency
Up to 16.2×
Lower is better
256-token prompts
1,000-token prompts
This runs on 4-year-old tech.
We will 5× the advantage.
Questions about the method or the results? Book a call with the founders, and we will run a benchmark on your own workload.
Book a call