1 points | by Phantasm0009 9 hours ago ago
1 comments
Hey HN! High school student here. Just hit v1.0.1 of accel-gpu - a WebGPU wrapper with NumPy-like API.
Went from 0 to 1,500+ downloads in 10 days with zero marketing. Something about "NumPy for browsers" resonates with people.
What's new in v1.0: • Cross-backend consistency tests (WebGPU/WebGL/CPU within 1e-4) • Memory management (tidy/scoped) with stress tests • Tree-shakeable imports (only load what you need) • Fixed critical WebGL matmul bug • Production-ready error messages
Why it exists: • Browser ML/AI is exploding (local-first, privacy) • But GPU libraries require shader knowledge (WGSL/GLSL) • TensorFlow.js is 10MB+ with steep learning curve • Wanted NumPy simplicity, ~160KB, zero dependencies
Try the demos: • Vector Search: https://phantasm0009.github.io/accel-gpu/example/vector-sear... (Search 100k docs client-side in <100ms) • Audio Visualizer: https://phantasm0009.github.io/accel-gpu/example/audio/ (Real-time FFT spectrogram) • Neural Network: https://phantasm0009.github.io/accel-gpu/example/nn/ (MNIST-style inference)
Tech deep-dive: • WebGPU → WebGL2 → CPU fallback (works everywhere) • Automatic shape inference • FinalizationRegistry for GPU cleanup • Playwright tests across Chromium/Firefox/WebKit • TypeScript with shape-level type safety
Built for college apps, but 1,500+ developers are using it. Feedback welcome!
Docs: https://phantasm0009.github.io/accel-gpu/
Hey HN! High school student here. Just hit v1.0.1 of accel-gpu - a WebGPU wrapper with NumPy-like API.
Went from 0 to 1,500+ downloads in 10 days with zero marketing. Something about "NumPy for browsers" resonates with people.
What's new in v1.0: • Cross-backend consistency tests (WebGPU/WebGL/CPU within 1e-4) • Memory management (tidy/scoped) with stress tests • Tree-shakeable imports (only load what you need) • Fixed critical WebGL matmul bug • Production-ready error messages
Why it exists: • Browser ML/AI is exploding (local-first, privacy) • But GPU libraries require shader knowledge (WGSL/GLSL) • TensorFlow.js is 10MB+ with steep learning curve • Wanted NumPy simplicity, ~160KB, zero dependencies
Try the demos: • Vector Search: https://phantasm0009.github.io/accel-gpu/example/vector-sear... (Search 100k docs client-side in <100ms) • Audio Visualizer: https://phantasm0009.github.io/accel-gpu/example/audio/ (Real-time FFT spectrogram) • Neural Network: https://phantasm0009.github.io/accel-gpu/example/nn/ (MNIST-style inference)
Tech deep-dive: • WebGPU → WebGL2 → CPU fallback (works everywhere) • Automatic shape inference • FinalizationRegistry for GPU cleanup • Playwright tests across Chromium/Firefox/WebKit • TypeScript with shape-level type safety
Built for college apps, but 1,500+ developers are using it. Feedback welcome!
Docs: https://phantasm0009.github.io/accel-gpu/