GitHub
Single, no cache
2.30B NPS
gigantua 1.6
CPU Leaderboard
Preview — 31 enginesA competition for the fastest move generator in the world. Engines submit a UCI-compatible binary, get validated against the standard perft corpus, and are ranked by throughput across four threading and cache configurations.
Last benchmarked:
Best in mode
Fastest engine in each of the four benchmark configurations. Throughput is the mean of best NPS across the six canonical positions — each timed at the deepest depth it can complete inside a fixed per-position budget.
Single, with cache
6.39B NPS
tgct_engine 0.1.15-r2r
Multi, no cache
33.2B NPS
perft_cpu_2026 main
Multi, with cache
361.2B NPS
perft_cpu_2026 main
Full results
One row per engine version. Bold cells mark the fastest engine in that mode. A dash (—) means the engine did not run that configuration. Click a column header to sort — missing values always sort last.
| Engine | Language | Version | Single, no cache | Single, with cache | Multi, no cache | Multi, with cache |
|---|---|---|---|---|---|---|
| gigantua | C++ | 1.6 | 2,304,487,198 (2.30B) | — | — | — |
| perft_cpu_2026 | C++ | main | 1,935,787,602 (1.94B) | 4,806,229,072 (4.81B) | 33,219,331,350 (33.2B) | 361,240,881,390 (361.2B) |
| chessbit | C++ | main | 1,895,360,906 (1.90B) | — | — | — |
| zerologic | C++ | 1 | 1,654,727,109 (1.65B) | 1,654,526,743 (1.65B) | — | — |
| mperft | C | 0be02ae | 1,541,075,983 (1.54B) | 4,751,986,393 (4.75B) | 26,822,399,079 (26.8B) | 171,925,899,207 (171.9B) |
| tgct_engine | C# | 0.1.15-r2r | 913,757,226 (913.8M) | 6,388,399,775 (6.39B) | 31,743,545,721 (31.7B) | 174,302,933,252 (174.3B) |
| caps | Rust | main | 789,354,941 (789.4M) | — | 17,394,610,319 (17.4B) | — |
| stockdory | C++ | cd237fc32a | 715,360,169 (715.4M) | — | 8,801,601,199 (8.80B) | — |
| clover | C++ | 9.1 | 652,228,686 (652.2M) | — | 647,103,120 (647.1M) | — |
| xiphos | C | master | 621,677,116 (621.7M) | — | — | — |
| cozy-chess | Rust | 0.3.4 | 602,090,829 (602.1M) | — | — | — |
| rose | C++ | main | 593,455,467 (593.5M) | — | — | — |
| potential | C | 3.40.88 | 566,946,992 (566.9M) | — | — | — |
| pawnocchio | Zig | 2.0-dev-a8f09e7 | 488,658,466 (488.7M) | — | — | — |
| raphael | C++ | 4.2.0-dev-87999df | 470,541,469 (470.5M) | — | — | — |
| chessnut | C++ | master | 463,630,235 (463.6M) | — | — | — |
| berserk | C | 13 | 461,515,307 (461.5M) | — | — | — |
| halogen | C++ | 13 | 431,256,530 (431.3M) | — | — | — |
| maestro | C++ | 1.2 | 401,516,637 (401.5M) | — | — | — |
| stash-bot | C | master | 378,347,944 (378.3M) | — | — | — |
| shakmaty | Rust | 0.27.3 | 370,826,239 (370.8M) | — | — | — |
| prune | C++ | dev | 352,422,736 (352.4M) | — | — | — |
| stormphrax | C++ | 7.0.108 | 352,063,841 (352.1M) | — | — | — |
| stockfish | C++ | 18 | 295,489,879 (295.5M) | — | — | — |
| jordanbray-chess | Rust | 3.2.0 | 284,905,153 (284.9M) | — | — | — |
| horsie | C++ | 1.1.7 | 254,493,846 (254.5M) | — | — | — |
| viridithas | Rust | 20.0.0-dev | 193,644,727 (193.6M) | — | — | — |
| quanticade | C | main | 124,070,057 (124.1M) | — | — | — |
| rubichess | C++ | 20260308 | 52,487,313 (52.5M) | — | 657,069,262 (657.1M) | — |
| ethereal | C | 14.40 (PEXT) | 46,946,368 (46.9M) | — | — | — |
| plentychess | C++ | 7.0.66 | 18,064,943 (18.1M) | — | — | — |
Drill down by run
Pick an engine, mode, and position to see how NPS scales with depth — and where subprocess startup amortizes into raw move-generation throughput.
Loading run data…
How it works
Three steps from binary to ranked entry. The full harness, position set, and verification rules live in the PerftWar directory.
1
Submit your engine
Any move generator that speaks the standard UCI go perft N protocol is eligible. Open the door on Discord to register and coordinate the upload.
2
Compliance check
Every entry runs through perftcheck against ~28,000 known-correct positions. Any wrong node count disqualifies the run — speed is the tiebreaker among engines that already pass.
3
Benchmark and rank
Validated engines are run on each of six TGCT-canonical positions, walking depths from d1 upward under a per-position wall-clock budget. We record the NPS at the deepest depth that completes, then mean those across positions for each of the four threading / cache configurations.
Caveats & methodology notes
A few things to keep in mind when reading these numbers:
-
Preview hardware
Until the bare-metal Ubuntu reference host is online, some entries are captured on developer hardware — look for version tags like -local. Expect absolute NPS to shift once the reference rig takes over.
-
Subprocess overhead
Each measurement spawns a fresh process so cache state is deterministic. At depth 4 the spawn cost can dominate — particularly for very small positions like pos3 (43k nodes). Depth 6 numbers are the more reliable cross-engine signal.
-
Partial mode coverage
Engines opt in to the modes they support. The reference engine currently exposes only cached configurations; a future build will fill in the no-cache modes once that switch lands.
-
Strict correctness first
Every subprocess's stdout is scanned for the expected node count. A single mismatch fails the whole engine run — no NPS is recorded for a disqualified entry.
Interested in entering?
The board is in preview with two engines. If you've got a move generator you'd like to put in the ring — in any language, any architecture, any style — drop into the project Discord and let us know.
The reference engine in MoveGen/ gives one starting point — but the leaderboard isn't tied to any one architecture. Magic bitboards, PEXT, classical ray-scan, 0×88 — all welcome.