Request the benchmark report

→Benchmark report

Request the measured numbers.

The full benchmark report behind the claims on the front page — methodology, hardware and software configuration, and the raw measured results. We send it to your work inbox, usually within one business day.

Name

Please tell us your name.

Work email

Please use your work email — we don't send the report to personal addresses (Gmail, Yahoo, Outlook, …).

Business address required. The report goes to this inbox.

Company

Please tell us where you work.

Role · optional

Your workload

Pick the closest match — "Other" is fine.

What are you serving? · optional

No list, no drip campaign — you get the report, that's it.

Something went wrong sending your request. Email us instead — hello@repercep.ai — and we'll reply with the report.

Request received●

Thanks — your request is in. The report will land in your work inbox, usually within one business day.

Wrong address or in a hurry? Email hello@repercep.ai.

In the report●

→
Cost per unit of work — H100 vs MI300X across every regime we measure, at current rental prices, net of the performance difference.
→
The serving levers — candidate batching, adaptive caching, session density: what each one is worth, on both vendors, with the caveats.
→
Parity gates — where we reproduce a reference implementation exactly, and where that check is still open. Stated either way.
→
Six models, three regimes — generation, energy-based control and policy-driven control, all on one unmodified engine.
→
Methodology — single GPU, end-to-end wall time, system-vs-system, no cluster asterisks.
→
Raw results — verbatim result records from every run, so you can check our arithmetic.

Measured · generation + closed-loop control · single GPU · Repercep runtime

We use your details only to send the report and follow up about your workload. Nothing is shared, nothing is sold, no mailing list.