tracel-ai/cubecl

Health v2 · maintenance only

Adoption not part of the score

2437 stars · 263 forks observed · 2026-10-16

What it is AI-extracted, prompt v1, taxonomy v1, 2026-08-30, confidence not recorded

CubeCL is a Rust language extension, JIT compiler, and runtime set for writing high-performance GPU compute kernels in plain Rust. A single #[cube] function compiles on demand to CUDA, HIP, Metal, SPIR-V, WGSL, or CPU SIMD with compile-time specialization and autotuning.

Use cases

• write gpu compute kernels in rust
• compile one kernel to cuda and webgpu
• use tensor cores from rust
• build a multi-backend linalg library
• avoid writing wgsl or cuda c by hand
• autotune kernel launch configurations

When to choose

• you need peak-performance kernels across nvidia, amd, apple, and webgpu backends
• you want type-checked, testable kernels in rust instead of shader strings
• you're building scientific computing or deep learning infrastructure in rust

When to avoid

• you just need simple graphics rendering rather than compute kernels
• you want a high-level deep learning framework rather than a kernel layer
• your project doesn't use rust

Facets

library · maturity active

compiler gpu-computing machine-learning math gpu-computing machine-learning programming-languages compilers performance rust cross-platform wasm jit-compiler cuda webgpu gpgpu compute-kernels tensor-cores autotuning rust-language-extension gpu

Member repositories

For agents

markdown · JSON · MCP: product_card(name="tracel-ai/cubecl")