cuda-wasm: CUDA to WebAssembly/WebGPU Transpiler

1.1.1 verified Fri May 01 auth: no javascript

cuda-wasm (v1.1.1) is a high-performance transpiler that converts CUDA C++ code to WebAssembly and WebGPU, enabling GPU-accelerated computing in browsers and Node.js without NVIDIA hardware. It uses a clean-room implementation based on public CUDA specifications, translating CUDA syntax to Rust/WebGPU patterns. Ships with TypeScript types, supports Node >=16.0.0. Key differentiators: no NVIDIA dependencies, near-native performance via WebGPU, ruv-FANN neural network integration. Released under MIT/Apache-2.0. Active development.

Common errors

error TypeError: WebGPU not supported in this browser ↓

cause Browser lacks WebGPU support (e.g., Safari, older Chrome).

fix

Use a browser with WebGPU enabled (Chrome 113+, Edge 113+, Firefox Nightly).

error Error: CUDA_RUNTIME_API_NOT_SUPPORTED ↓

cause Using CUDA runtime API calls (e.g., cudaMalloc) in the transpiled code.

fix

Rewrite kernel to avoid runtime API; allocate memory outside and pass as args.

Warnings

breaking CUDA runtime API and driver API are NOT supported. Only kernel execution is transpiled. ↓

fix Use transpileCUDA and runKernel for kernels; do not expect CUDA runtime functions.

breaking Not all CUDA features are implemented. See compatibility table. ↓

fix Consult the documentation for supported features; avoid dynamic parallelism and PTX.

gotcha WebGPU requires HTTPS or localhost. Using file:// or non-secure origins will fail. ↓

fix Serve your app over HTTPS or use localhost; do not open HTML directly from filesystem.

deprecated The separate @cuda-wasm/web package is deprecated and merged into main package. ↓

fix Use cuda-wasm directly; no need for additional packages.

Install

npm install cuda-wasm

yarn add cuda-wasm

pnpm add cuda-wasm

Imports

cudaWasm

wrong

const cudaWasm = require('cuda-wasm')

correct

import { cudaWasm } from 'cuda-wasm'

Package is ESM-only; CommonJS require not supported.

transpileCUDA

wrong

import { transpileCUDA } from 'cuda-wasm'

correct

import { transpileCUDA } from 'cuda-wasm/transpile'

Subpath export: use '/transpile' for the transpiler function.

runKernel

wrong

import { runKernel } from 'cuda-wasm'

correct

import { runKernel } from 'cuda-wasm/runtime'

Runtime execution is in '/runtime' subpath.

NeuralNetwork

wrong

import { NeuralNetwork } from 'cuda-wasm/neural-network'

correct

import { NeuralNetwork } from 'cuda-wasm/neural'

Neural network module is at '/neural'.

Quickstart

Transpile a CUDA kernel and run it on WebGPU, demonstrating vector addition.

import { transpileCUDA } from 'cuda-wasm/transpile';
import { runKernel } from 'cuda-wasm/runtime';

const cudaCode = `
__global__ void vectorAdd(const float* A, const float* B, float* C, int N) {
  int i = threadIdx.x + blockIdx.x * blockDim.x;
  if (i < N) C[i] = A[i] + B[i];
}
`;

async function main() {
  const wasmModule = await transpileCUDA(cudaCode);
  const result = await runKernel(wasmModule, 'vectorAdd', {
    args: [new Float32Array([1,2,3]), new Float32Array([4,5,6]), new Float32Array(3), 3],
    grid: [1, 1, 1],
    block: [3, 1, 1]
  });
  console.log(result); // Float32Array [5, 7, 9]
}
main();