Benchmarked against open-source optimisation
The same 10-year open-pit scenario on a model of approximately 70,000 blocks, the same variables, constraints and assumptions, given to Google OR-Tools and to Cognomine. Here is what came back.
| Metric | Google OR-Tools | Cognomine | Cognomine advantage |
|---|---|---|---|
| Discounted NPV | $1,667M | $1,940M | +$273M (~16%) |
| Equipment capital | $254M | $104.5M | $149.5M lower (~59% less) |
| Peak fleet | 5 diggers / 41 trucks | 2 diggers / 17 trucks | Fewer than half the diggers; 24 fewer trucks |
| Time to produce the plan | 12 h+ ¹ | ~25 seconds ² | More than 1,000x faster ¹ |
- The OR-Tools run was stopped unconverged after 12 hours. The speed multiple is a lower bound.
- Solver-to-solver comparison. OR-Tools is a CPU-only solver; Cognomine ran on cloud GPUs, the hardware it is built for.
- Both runs used an identical scenario: the same variable count, constraint sets and assumption sets.
- The extraction boundary was taken from the input data for both solvers, which were each free to leave blocks unmined. Neither run solved for the ultimate pit.
What the numbers mean
On this scenario, Cognomine found a plan worth $273M more in discounted NPV than the OR-Tools result, while cutting equipment capital by $149.5M. The gain comes from better sequencing: the plan reaches similar production with a much smaller peak fleet.
Where OR-Tools was stopped after 12 hours without converging, Cognomine returned its plan in around 25 seconds. That difference changes how planning works. Instead of committing to one run, a planner can test dozens of assumptions in an afternoon.
Benchmark results come from a single test scenario and outcomes depend on your deposit, constraints and assumptions. The comparison uses an open-source baseline because it is reproducible and free of licensing restrictions. Cognomine produces high-quality feasible plans.
Run it on your own numbers
We also publish a benchmark on public data measuring whether plans that score well can actually be executed. See the MineLib benchmark