IdeaWell

Project ideas from Hacker News discussions.

Numba in the Browser: Unlocking a New Scientific Python Stack in JupyterLite

Original Article

Hacker News Discussion

📝 Discussion Summary (Click to expand)

Three dominant themes in the discussion

  1. Broad enthusiasm for the breakthrough

    “Congrats for a serious engineering achievement!” – rossant

  2. Concern over the practical overhead

    “Numba+llvmlite alone is almost 100MB!” – momojo

  3. Interest in future integrations and prior work

    “It also works with Pytensor & PyMC!” – SylvainCorlay

🚀 Project Ideas

MicroNumba Browser Runtime

Summary

  • Enables on‑demand JIT compilation of tiny NumPy‑style kernels in the browser without pulling in the full 100 MB Numba stack.
  • Solves the excessive bundle size problem highlighted by momojo.

Details

Key Value
Target Audience Scientists and data‑engineers who need lightweight NumPy‑like performance in web apps.
Core Feature Tiny WebAssembly JIT that compiles only the required kernels, cutting runtime size by >90 %.
Tech Stack Rust + wasm‑bindgen, LLVM‑lite, optional Emscripten, bundled with esbuild.
Difficulty High
Monetization Revenue-ready: SaaS hosting with per‑kernel compilation credits.

Notes

  • HN commenters would love it because “it fixes the 100 MB bloat” and “makes in‑browser scientific computing truly local‑first”.
  • Could spark discussion about alternative JIT strategies and attract open‑source contributors.

BenchBrowser: In‑Browser Performance Benchmark Suite

Summary

  • Provides a ready‑to‑use UI for comparing absolute performance numbers of code snippets between native execution and WebAssembly/WASM‑compiled versions.
  • Directly answers lmc’s request for concrete benchmark figures.

Details

Key Value
Target Audience Developers evaluating browser‑based compute vs native performance.
Core Feature Interactive benchmark runner that reports wall‑clock times, throughput, and speed‑up ratios for any JS/Numba‑style function.
Tech Stack TypeScript front‑end, Web Workers for isolation, Node‑based harness for native baseline, Plotly for results.
Difficulty Medium
Monetization Hobby

Notes

  • HN users would love the ability to “see the 250× vs 90× claim in real numbers” and discuss optimization tricks.
  • Potential for community‑contributed benchmark suites and a repository of performance regressions.

JAX‑WebGPU Playground

Summary

  • A minimal, browser‑native environment that compiles JAX operations to WebGPU shaders, allowing early experimentation with JAX‑style autodiff and JIT in the browser.
  • Addresses hessammehr’s call for “jax in the browser on WebGPU”.

Details

Key Value
Target Audience Machine‑learning researchers and functional programming enthusiasts.
Core Feature Transpiler that converts JAX XLA graphs to WebGPU compute pipelines, exposing gradient‑aware tensors in‑browser.
Tech Stack Python 3.11 (for prototyping), spaCy‑based AST parser, glsl‑canvas, Tailwind UI, optional Pyodide bridge.
Difficulty High
Monetization Hobby

Notes

  • Commenters would love “a first‑step toward JAX on the web” and would likely discuss performance trade‑offs and educational use‑cases.
  • Could evolve into a collaborative playground, fostering discussion on WebGPU’s suitability for scientific computing.

Read Later