Show HN: RunMat – runtime with auto CPU/GPU routing for dense math
9 days ago
- #Open Source
- #GPU Acceleration
- #MATLAB Syntax
- RunMat automatically fuses operations and intelligently routes between CPU and GPU using MATLAB syntax without requiring kernel code or rewrites.
- RunMat is in early build with core runtime and GPU engine passing thousands of tests, though some plotting features are missing or buggy.
- It supports MATLAB-style syntax, automatically optimizing operations for CPU or GPU based on data size and transfer cost heuristics.
- RunMat works across various GPUs (NVIDIA, AMD, Apple Silicon, Intel) via native APIs like Metal, DirectX 12, and Vulkan, ensuring no vendor lock-in.
- Features include a modern CPU runtime with an interpreter for fast startup, a JIT for hot paths, and a memory-safe design using Rust.
- The GPU backend uses wgpu/WebGPU, supporting multiple platforms and falling back to CPU for small workloads.
- Plotting and tooling are in pre-release, with basic 2D plots functional and more advanced features like 3D plots on the roadmap.
- RunMat is open-source under the MIT License, emphasizing a small binary, CLI-first design, and community-driven package ecosystem.
- Benchmarks show RunMat outperforming PyTorch and NumPy in various scenarios, especially on large arrays.
- Installation is straightforward via shell scripts, crates.io, or building from source, with support for interactive REPL, script execution, and Jupyter kernels.