Efficiently updatable neural network

This article is part of the series on
Chess programming
Board representations 0x88 Bitboards
Evaluation functions Deep neural networks (Transformers) Attention Efficiently updatable neural networks Handcrafted evaluation functions Piece-square tables Reinforcement learning Stochastic gradient descent Supervised learning Texel tuning Unsupervised learning
Graph and tree search algorithms Minimax Alpha-beta pruning Principal variation search Quiescence search Monte Carlo tree search
Chess computers Belle ChessMachine ChipTest Cray Blitz Deep Blue Deep Thought HiTech Hydra Mephisto Saitek
Chess engines AlphaZero Chess Tiger Crafty CuckooChess Deep Fritz Dragon by Komodo Chess Fairy-Max Fritz Fruit GNU Chess HIARCS Houdini Ikarus Junior KnightCap Komodo Leela Chess Zero MChess Pro Mittens MuZero Naum REBEL Rybka Shredder Sjeng SmarThink Stockfish Torch Turochamp Zappa
v t e

In computer strategy games like in shogi and chess, an efficiently updatable neural network (NNUE, a Japanese wordplay on Nue, sometimes stylised as ƎUИИ) is a neural network-based evaluation function whose inputs are piece-square tables, or variants thereof like the king-piece-square table.^[1] NNUE relies on the tendency in these games for the game state to change only slightly between moves.

NNUE was invented by Yu Nasu and introduced to computer shogi in 2018.^[2][3] On 6 August 2020, NNUE was for the first time ported to a chess engine, Stockfish 12, resulting in a major increase in playing strength for that engine.^[4][5] As of 2025, NNUE is the dominant evaluation function paradigm used by top alpha–beta chess engines like Komodo Dragon.

NNUE are designed to run efficiently on central processing units (CPU).^[6][7] They use used incremental computation and single instruction multiple data (SIMD) techniques along with appropriate intrinsic instructions.^[2] In contrast, deep neural network-based chess engines such as Leela Chess Zero rely on without a requirement for a graphics processing unit GPUs for efficient inference.^[8][9]

The neural network used for the original 2018 computer shogi implementation consists of four weight layers: W1 (16-bit integers) and W2, W3 and W4 (8-bit). It has 4 fully-connected layers, ReLU activation functions, and outputs a single number, being the score of the board. As of 2025, Stockfish has introduced several optimizations to the NNUE architecture, but the overall architecture remains similar.^[10]