-
nocturn9x released this
2026-06-27 16:01:35 +02:00 | 4 commits to master since this releaseMore layers!
Ahoy! As it was foretold by the prophecy, I come to you today with a new major release of Heimdall, featuring a multi-layer NNUE for much-improved evaluation quality. Exciting!
What? You're asking me why I am releasing 1.5 so close to the release of 1.4.3? What are you talking about? There is no Heimdall 1.4.3, and any references or mentions to it are pure fiction. Seriously. Stop talking about it. (Seriously. You don't want to upset the Elo gods, do you?)
Anyway, here follows a changelog
What changed
- Added a (shamelessly vibecoded) terminal UI with analysis, play/setup/replay flows, board rendering, eval/move arrows, user arrows, mate finding, annotations, PGN importing/exporting, and more; see here for details
- Made UCI move parsing reject illegal moves more reliably
- Improved command parsing so valid
positionandgocommands embedded in otherwise malformed input can still be handled per the UCI spec - Switched to the first multilayer network in Heimdall's history,
gramr - Added a new
vnnitarget for improved inference speed on AVX512 processors that support it (no AVX2-VNNI, sorry Intel people) - Added Transparent Huge Pages support for aligned memory where available (detected at build time, not runtime)
- Implemented a simplified ProbCut
- Implemented fractional LMR and made the LMR base table tunable
- Added continuation correction history and 2-ply continuation correction history
- Tuned all search parameters
- Fixed evaluation clamping and mate-score handling issues
- Fixed the long-standing incorrect definition of
improving - Tightened singular-extension and futility-pruning margins
- Ensured NMP does not return false mates
- Reduced false sharing and contention by changing how search parameters and worker state are copied/aligned
- Used relaxed ordering for atomic operations where appropriate
- Moved more objects to managed references and aligned more search-manager fields
- Compressed moves into 16 bits
- Heimdall is now NUMA-aware! On Linux machines with more than one processor (if configured with more than one thread), Heimdall will attempt to bind the transposition table and search threads in a way that maximizes performance and prevents needless back-and-forth between the different memory controllers. Should be significant at TCEC/CCC conditions
- Reduced the time it takes for many-thread searches to start. Relevant at very high thread counts when low on time
- En Passant moves are now considered as captures across search and movegen rather than being special-cased
- IDK, probably forgot a bunch of stuff, but who ever reads changelogs anyway?
Strength
The keyword for this release is definitely scaling. That fat ass network surely seems to at least be worth the weight! The 8moves tests are benefitting 1.4.2 as well so the scaling is less pronounced, but I'm confident it's still there
UHO_Lichess_4852_v1.epd
STC:
Elo | 82.77 +- 9.57 (95%) Conf | 8.0+0.08s Threads=1 Hash=16MB Games | N: 2010 W: 800 L: 330 D: 880 Penta | [13, 106, 365, 440, 81]https://chess.n9x.co/test/7152/
LTC:
Elo | 97.81 +- 8.43 (95%) Conf | 40.0+0.40s Threads=1 Hash=64MB Games | N: 2016 W: 832 L: 279 D: 905 Penta | [2, 76, 352, 523, 55]https://chess.n9x.co/test/7154/
8moves_v3.epd
STC:
Elo | 80.66 +- 9.06 (95%) Conf | 8.0+0.08s Threads=1 Hash=16MB Games | N: 2030 W: 644 L: 181 D: 1205 Penta | [9, 90, 425, 411, 80]https://chess.n9x.co/test/7156/
LTC:
Elo | 71.80 +- 7.34 (95%) Conf | 40.0+0.40s Threads=1 Hash=64MB Games | N: 2012 W: 491 L: 81 D: 1440 Penta | [2, 47, 535, 383, 39]Downloads