010 · 22 October 2026 · developer-tools · machine-learning · education

Kiln Workbench

Download ↓kiln-workbench-0.1.0.html, 725 KB · sign in to download

A compiler workbench: write passes and IRs in JavaScript, watch every rewrite, generate CUDA and WGSL, and measure kernels on simulated GPUs.

Previewin your browser
The attention lesson: notes with the maths next to the kernel graph and measurements
Sign in to run
Download ↓kiln-workbench-0.1.0.html, 725 KB · sign in to download

A compiler workbench. Write programs, passes and IRs in JavaScript, watch every rewrite, generate CUDA and WGSL, and measure kernels on simulated GPUs, on WebGPU, or on a CUDA box.

  • Lessons with live code: presets are short chapters with the maths, the program, what you should see, and exercises: the roofline, why the compiler fuses, the matmul ladder, occupancy and latency hiding, one program on five GPUs.
  • Real kernels: RMSNorm fused, softmax in one kernel, a tiled linear layer, attention, N-body forces, a box-blur stencil, and a regex compiled to an NFA, a DFA and JavaScript.
  • Extend it: write your own pass, your own cost model or a new frontend op.
  • Every level visible: from the torch-like graph down through tensor, loop, tile and kernel IRs to GPU code, with a trace of each rewrite and the rule that made it.
  • Measure: knobs and sliders feed the kernel; simulated devices report Nsight-style metrics and compare against the reference output.
  • Agent: a coding agent that can read, change and test the program alongside you.