hpc-labs.ai

AI-boosted optimization, from code to silicon.

Consulting services

Not every project needs a licence. We take on sporadic projects, and we plan paid benchmarking reports for questions that cost real money and time to answer. Both use the same measured, verified method as ParetoStack: every change is timed against the original and kept only when the output is verified.

Sporadic projects

Optimize a code base. Python, Java, C and C++: build settings, loops, libraries, data layout, each change verified against your own output. You get the changes as a patch and a report.
Cloud function cost and speed. AWS Lambda and Google Cloud Functions, measured in throwaway copies in your own account.
Port or tune for a target. Arm, x86, GPU: measured on the hardware you name.
Hardware design flows. Verilog and SystemVerilog on open-source tools (in development): synthesis recipes and place-and-route spread, checked by simulation.

Quoted per project after a short call; the quote says what is measured, on which machine, and what is not promised. We do not promise a speed-up: some code has none to give, and the report says so.

Paid benchmarking reports (planned)

Reports for a decision that is expensive to test yourself. Each report is a PDF with the method, the machines, the raw numbers and what the numbers do not show. It is licensed to one buyer and watermarked with the buyer's name. None is on sale yet; the first is in preparation.

Which AI engine optimizes a large code base best? · first report, OpenCV 5.x

The same optimization tasks on OpenCV 5.x, given to different AI engines, with every proposed change checked against the original output and timed on real hardware:

Engines compared: OpenAI Codex, Claude Code with Claude Fable 5, and an open-weight model run locally (from a Chinese lab).
What you learn: the verified speed-up each engine found, how many of its changes failed the check, and the tokens, time and cost per verified gain. A local model costs no tokens but needs hardware: the report prices both.
Method: ParetoStack's measured loop and output contract; one machine and one OpenCV version per table, stated in the report.
Planned price: US$2,999 per report, one buyer licence, watermarked. Not for sale yet: write to us to be told when it is ready, or to ask for a topic.

The results are not known yet, and the report will publish whatever is measured, including an engine that does badly.

More topics are possible; tell us what you would pay to know.