Every part testable on its own.
HexaGrid is a scheduler, not a suite. Live grid prices come in, a demand forecast says where your load is going, a constraint solver decides what runs when, and a test harness checks the whole thing against cases where the right answer is already known. This page is the detail behind that.
Live grid and carbon feeds
HexaGrid integrates wholesale price data for all five US ISO regions through the EIA API, cached with a 15-minute TTL, plus carbon intensity for the same regions. In a deployment these are the live inputs every scheduling decision is built on.
What your site will draw
A neural network predicts facility demand up to two hours ahead, so deferral decisions are made against where load is heading rather than where it just was. It is measured against the honest baseline — assuming demand stays where it is now — on data the model has never seen.
| Horizon | HexaGrid error | Last-value baseline | Better by |
|---|---|---|---|
| 30 min | 0.60 kW | 2.9 kW | 5× |
| 60 min | 0.56 kW | 5.2 kW | 9× |
| 120 min | 0.59 kW | 7.3 kW | 12× |
The point is the shape rather than the ratio. HexaGrid's error stays flat as the horizon grows, while the baseline degrades steadily — which is what makes a two-hour deferral window usable at all.
A constraint solver, not a heuristic
Jobs, cluster capacity and deadlines are handed to CP-SAT as a real
scheduling model — intervals, cumulative capacity constraints and element lookups for
time-varying prices. That formulation is what keeps it tractable across a 300-run sweep.
Every solve returns a certified bound: not just a schedule, but a proof of how far it could possibly be from the best one. Across 300 solves the median certified gap was 0.32% and the worst was 3.50%, with 102 proved exactly optimal. You are told how good the answer is rather than asked to assume it.
Three arms, including a ceiling we cannot beat
Every benchmark runs three ways on the same jobs and the same real price data. The middle arm is the product. The outer two exist so you can see how much room there is on either side of it.
On arrival
Every job runs the moment it lands. No scheduling. This is the baseline the 13.8% is measured against.
HexaGrid
Scheduled against the forecast, then billed at the prices that actually occurred. Forecast error costs you real money here, as it would in production.
Perfect foresight
Scheduled with tomorrow's prices known in advance. Impossible to build. On this data it saves 16.0% — the hard ceiling for the whole idea, and we capture 88% of it.
A trade you make, not a score we invent
The same schedule can be tuned toward cost or toward emissions. HexaGrid solves for a point on that frontier and reports both outcomes in their own units — dollars and tonnes — and never adds them into a single number, because the exchange rate between them is yours to set, not ours.
Tests that are allowed to fail
Nineteen regression tests run on every build. Several are null cases where the correct answer is known in advance — given a flat price curve the scheduler must report exactly zero saving, savings must increase with price volatility, and zero deferrable work must yield zero benefit. An earlier version of this code reported 14.8% savings on a flat curve. The current one reports exactly 0.00% across all 50 such runs, and the test that catches it runs on every build.
On the hardware side, NVML telemetry polls every GPU every ten seconds — temperature, power draw, memory use, ECC counts, fan speed — and an anomaly detector watches for combined-metric failure signatures that no single threshold would catch. Nothing gets scheduled onto hardware that is struggling.
Runs beside your cluster
No agents on your nodes, no hardware changes, no vendor lock-in. A single service process, a dashboard, and two API keys. Nothing to evaluate costs you anything.