Notes
Technical Notes
Each note reports a measurement and the conditions it was taken under, so an engineer can check the method before talking to us.
Series
| HS‑002 |
Making PyTorch skip the part that didn’t change A PyTorch extension for loops that rebuild the same large matrix product every step when almost nothing has moved. 16.7× against the best incremental version you can write without leaving PyTorch — though most of that is the cost of the PyTorch call itself, and the note shows the breakdown. Against a hand-written C++ equivalent it runs 6.9× down to 4.8× while the changed set moves gradually, 2.9× to 1.4× while it moves at random. |
2026-08-07 |
|---|---|---|
| HS‑001 |
Raising the processing ceiling of a Kafka Streams aggregation
benchmark A standard published streaming benchmark, re-run on x86 with our work in place. 15× the rate at which it drained a standing backlog, with lower processor load and a smaller memory footprint. Five of the six things that changed between the two runs are architecture rather than our method, and we did not run the arm that would separate them. |
2026-07-18 |
Raw measurement data for our published benchmark sweeps is available under /benchmarks/.
The same measurement, on your workload.
An assessment runs your system on your own hardware and reports the difference. It starts with a call.
Book a discovery call →