performance.md
July 26, 2026 · View on GitHub
🚀 How We Made the Solver 2 to 3 Times Faster
The solver is now 2 to 3 times faster per finished frame than the previous version (commit 97ac227, May 21, 2026). Running the same scenes on the same machine, every scene got faster: on average about 2.9x per finished frame, with a median of 2.2x, and up to 5.4x on the scenes that benefit most.
🧪 Method
- Same scenes and settings for both versions; only the solver differed.
- One NVIDIA L40S, scenes run one at a time so nothing competed for the card.
- "Per frame" is the wall-clock time to finish one frame of the animation. This is the number that matters, because it is what you actually wait for.
📊 Results
| Scene | Faster per frame |
|---|---|
| trampoline | 5.40x |
| drape | 5.16x |
| cards | 4.82x |
| friction | 3.75x |
| belt | 3.65x |
| hang | 3.62x |
| yarn | 3.23x |
| domino | 3.17x |
| roller | 2.55x |
| woven | 2.50x |
| fishingknot | 2.23x |
| stack | 2.23x |
| needle | 2.14x |
| curtain | 2.13x |
| fitting | 2.11x |
| ribbon | 2.06x |
| twist | 2.03x |
| five-twist | 2.01x |
| trapped | 1.94x |
| noodle | 1.91x |
| codim | 1.73x |
Average: about 2.9x per finished frame, with a median of 2.2x. Nothing got slower.