Lc0: Open-Source Neural Network Chess Engine and GPU Acceleration in Practice

Lc0 is the core engine of the Leela Chess Zero project, built on deep neural networks and Monte Carlo Tree Search (MCTS) technology. It complies with the UCI protocol and supports cross-platform GPU acceleration and wide hardware compatibility. As an open-source continuation of the AlphaZero technical route, it represents a significant breakthrough for AI in the gaming domain. Driven purely by neural networks for decision-making, it provides a powerful tool for chess enthusiasts and AI researchers.

The Leela Chess Zero project started in 2018 with the goal of reproducing DeepMind AlphaZero's achievements in chess and making it open-source. Unlike traditional engines like Stockfish that rely on manual evaluation functions, Lc0 is purely driven by neural networks and learns from scratch through self-play. Its core engine complies with the UCI protocol and can be used with graphical interfaces such as Arena and Fritz.

The core of Lc0 is a deep residual convolutional network, including a policy head (predicting move probabilities) and a value head (evaluating the probability of winning/losing a position). Training is done through self-play: moves are selected using MCTS, and network weights are updated after the game ends. The MCTS process includes selection (PUCT algorithm), expansion (neural network evaluation and expansion of legal moves), backup (backpropagation of evaluation results), and repetition until the computational budget is exhausted.

Lc0 supports multiple computing backends:

• NVIDIA GPU: CUDA/cuDNN supports Maxwell to Ada Lovelace architectures; Tensor Core GPUs can use mixed-precision acceleration;
• AMD/Intel GPU: Experimental SYCL support since v0.32.0 (requires Intel oneAPI);
• CPU: Compatible with OpenBLAS/DNNL;
• ONNX Runtime: Includes backends like onnx-cpu/cuda/trt/rocm/dml;
• Apple Metal: Supports Apple Silicon and Intel Macs.

Lc0 uses Meson for building, requiring a C++20 compiler, Meson/Ninja, Git, and corresponding backend libraries. Official precompiled binaries are provided. To run it, you need to download neural network weight files (obtained from the official website; they come in different sizes—larger networks have stronger chess strength but higher resource requirements).

Lc0 ranks among the top in CCRL ratings, on par with Stockfish. Its playing style features:

• Focus on long-term positional advantages, sacrificing short-term material gains;
• Tactical ability comparable to traditional engines when given sufficient thinking time;
• Adjustable parameters to simulate different styles. Performance depends on hardware; high-end GPUs (e.g., RTX4090) can achieve tens of thousands of evaluations per second.

Lc0's application scenarios include:

• Game analysis: High-accuracy positional evaluation helps study complex positions;
• Practice games: Adjustable difficulty suitable for users of different levels;
• Opening research: Discover new ideas beyond traditional theory;
• Engine battles: Duels with Stockfish and others serve as benchmark tests for AI algorithms and hardware.

Lc0 is an active open-source project with development on GitHub, where the community contributes code and discussions. Training resources come from distributed training with volunteer GPU donations. It is not only a powerful engine but also a model of deep learning in game AI. In the future, it will push the boundaries of chess AI with hardware and algorithm optimizations.