PulseGTO
Benchmarks & research
Performance & memory
Preflop engine memory
Configuration: a production-size training state with 84–97 million information sets, i5-1235U, 64 GB installed RAM.
| Measurement | Value |
|---|---|
| Working set during training, production-size state | 16.6 GB |
| Time per sample, same state | about 3 s |
| Peak memory, small test run with flop enumeration | 3.0 GB |
Performance is measured on a resumed production-size checkpoint, because a small test run does not show every cost.
RangeCFR memory profiles
Configuration: heads-up RangeCFR audit trees. The engine's own RAM estimate matched the measured RSS to within 0.09%.
| Profile | RAM vs fp32 | Effect on the result |
|---|---|---|
| fp32 | 100% | Reference |
| strategy16 (default) | ≈75% | Kernels unchanged |
| compact16 | ≈50% | Root frequencies moved by up to 1.6 pts |
Heads-up RangeCFR audit trees. compact16 moved root frequencies by up to 1.6 points.
Three-way showdown and re-solve
See the payoff validation and the desktop timings.
A note on laptops
Windows can throttle a program started on battery power and move it to the efficiency cores; on the test laptop that meant 27 s per sample instead of 3. The preflop engine opts out of Windows power throttling when it starts. If a run is much slower than expected, check the power plan first.