AlphaTensor, developed by Google DeepMind, is the research codebase accompanying the 2022 Nature publication “Discovering faster matrix multiplication algorithms with reinforcement learning.” The project demonstrates how reinforcement learning can be used to automatically discover efficient algorithms for matrix multiplication — a fundamental operation in computer science and numerical computation. The repository is organized into four main components: algorithms, benchmarking, nonequivalence, and recombination. These contain implementations of the discovered matrix multiplication algorithms, tools to benchmark their real-world performance, proofs of nonequivalence among thousands of solutions, and methods for decomposing larger problems into smaller factorizations. Users can explore AlphaTensor’s discovered algorithms interactively using Colab notebooks or Python scripts.
Features
- Implements algorithms for matrix multiplication discovered via reinforcement learning
- Includes 14,000+ nonequivalent algorithms for 4×4 matrix multiplication problems
- Provides Colab notebooks for interactive exploration and verification
- Benchmarking scripts for performance evaluation on NVIDIA GPUs
- Python tools for tensor recombination and algorithm decomposition
- Reproducible experiments corresponding to the Nature 2022 publication