What we do
We make existing systems faster without new hardware.
The work we remove is nearly always the same shape. A system holds a structure that does not change — a grid, a network, a fixed set of relationships — and rebuilds every result over it on every cycle, even though only a small part of the input moved.
Removing it is mostly a question of how data is stored and what gets reused, not of exotic algorithms. On a dense operator, changing the storage layout alone was worth 3.2–8.1× over a delta implementation that was already skipping the unchanged inputs.
We work in-process, in your language, against your own workload. There is no service to call and nothing to send us.
How we know
We measure first, and we say when it will not help.
Every number we publish carries what it was measured against, on which machine, with which compiler, and which control ran. Ratios without those are not results, and we have withdrawn our own when they turned out to be measuring the baseline rather than the change.
The stream running above is one of our own benchmark captures, and the tiles are diffed live from the frame rather than replayed. On a fixed camera the scene barely moves, which is the whole opportunity. Read both rungs:
One method is narrow. The work is not. The technique behind the figures below fits some steps of a computation and not others, and where it does not fit it is slower than leaving your system alone: roughly half the speed of a good existing implementation when nothing moves between cycles. That is expected behaviour rather than a defect, and we would rather tell you before an invoice than after one. Where it fits, and where it does not.
- Against rebuilding the field every step
- 47×
- Against an implementation already skipping unchanged cells
- 1.29×
- Difference from the full recomputation
- 3.3×10−17
Raw measurement files: benchmarks/.
1000×1000 synthetic diffusion field, one moving source, about 2% of cells changing per step; MacBook M1. The first number is the one most vendors quote and the second is the one that tells you whether we are worth hiring. If your system already does the second thing well, we will say so.
Start
Start with a number for your own system.
A discovery call is thirty minutes, and the four questions we ask fit in a paragraph: what the system does, how often it recomputes and how long that takes today, what it runs on, and what the result has to be exact to. No preparation, and we do not need to see any code.
From there an assessment runs your own workload on the hardware you use today and reports what it could run at. The report is short: the number, the conditions it was taken under, and the specific changes in your system that would get you there. Two to four weeks, fixed fee, quoted before we start. Everything runs on your machines and nothing is transmitted to us.
Book a discovery call →