December 2, 2025
Fast math, identity crisis
Show HN: RunMat – runtime with auto CPU/GPU routing for dense math
Blazing-fast math app sparks a “who is this for?” brawl
TLDR: RunMat promises auto CPU/GPU speed with MATLAB‑style code, open source and cross‑device, but it’s still pre‑release. The comment vibe fixates on one question—who’s this for?—splitting hype over speed claims and freedom from lock‑in against skepticism about benchmarks and missing features.
RunMat just crash‑landed into Hacker News with a bold claim: write MATLAB‑style code and it will automatically decide whether your math runs on the computer’s brain (CPU) or the graphics chip that’s great at crunching numbers (GPU). The devs flaunt big speed charts and cross‑device support (NVIDIA, AMD, Apple, Intel), plus an open MIT license and a simple, code‑first vibe. But the crowd reaction boiled down to one spicy mood: identity crisis. As one top voice put it, who is this actually for? MATLAB folks escaping pricey licenses? Python people who live in notebooks? Julia diehards who already brag about speed? That question set the tone.
Fans cheered the promise of “no vendor lock‑in” and loved the idea of writing familiar scripts that run anywhere. Skeptics side‑eyed the benchmarks—mostly on Apple’s M2 Max—asking for real projects, not lab demos. The pre‑release status (v0.2) with buggy/missing plots also nudged people into “cool, but call me when it’s ready.” The memes flew: nostalgia for MATLAB, Pythonistas clutching their package manager, Julia fans whispering “we’ve been here.” The real drama isn’t the speed graphs—it’s the turf war over audience and workflow. Want to pick a side? Peek the RunMat site and bring popcorn.
Key Points
- •RunMat (v0.2) is a MATLAB-style runtime that automatically fuses operations and routes execution between CPU and GPU based on heuristics.
- •It supports GPUs from NVIDIA, AMD, Apple Silicon, and Intel via Metal, DirectX 12, and Vulkan using a wgpu/WebGPU backend.
- •The CPU runtime uses an Ignition interpreter, a Turbine JIT (Cranelift), and a generational GC, with Rust-based memory safety.
- •Benchmarks on Apple M2 Max show significant speedups over NumPy and PyTorch for large workloads; small tasks may stay on CPU.
- •RunMat is open source under an MIT License; plotting is limited in pre-release, and installation is available via scripts, cargo, or source.