# chelsea0x3b/cudarc

Safe rust wrapper around CUDA toolkit

Repository: https://github.com/chelsea0x3b/cudarc
Canonical: https://ross.abutalabs.com/products/cudarc
Language: Rust
License: Apache-2.0
License Family: permissive
Topics: cuda, cuda-programming, cuda-toolkit, gpu, gpu-acceleration, rust, cublas, curand, cuda-kernels, nvrtc, cudnn, nccl
Last push: 2026-08-12T14:37:42+00:00

## Health v2 (maintenance only)
Score: 98/100 (v2, computed 2026-09-03T02:20:16.233290+00:00)
- activity 97, release rhythm 97, longevity 100
- inputs: {"age_days": 1447, "days_push": 21, "days_rel": 22, "gap_med": 8, "n_releases_24m": 44}
- flags: none
- formula: round(0.45*activity + 0.35*rhythm + 0.20*longevity); archived -> min(score, 10)

## Adoption (not part of the score)
Stars 1213, forks 166 (observed 2026-08-28T04:04:00.476344+00:00)

## What it is
cudarc is a safe, minimal Rust wrapper around the NVIDIA CUDA toolkit, exposing the CUDA driver API plus libraries such as NVRTC, cuBLAS/cuBLASLt, cuDNN, cuRAND, NCCL, cuSPARSE, cuSOLVER, cuFFT, cuFILE, and CUPTI. It supports dynamic loading (no build-time CUDA required), dynamic linking, and static linking, with cargo feature flags to select CUDA 11.4-13.3, cuDNN 8.9-9.21, and NCCL 2.22-2.30 versions.

## Use cases
- run custom CUDA kernels from Rust
- call cuBLAS or cuDNN from a Rust program
- safely allocate and transfer GPU memory in Rust
- compile CUDA kernels at runtime with NVRTC
- do multi-GPU communication with NCCL in Rust
- GPU-accelerate a Rust application on NVIDIA hardware
- use CUDA without a CUDA toolkit installed at build time
- write deep learning primitives in Rust on top of cuDNN

## When to choose
- you are writing Rust that needs NVIDIA GPU acceleration
- you want safe, idiomatic handling of CUDA contexts, streams, and device memory
- you need FFI bindings to cuBLAS, cuDNN, cuRAND, or NCCL without writing unsafe C interop yourself
- you want to build without a local CUDA toolkit via the dynamic-loading feature
- you need to pin bindings to a specific CUDA, cuDNN, or NCCL version

## When to avoid
- you target AMD or Intel GPUs rather than NVIDIA CUDA
- you want a high-level tensor library with autodiff rather than raw CUDA bindings
- you are not working in Rust
- you need exhaustive coverage of every CUDA API, since cudarc intentionally exposes a minimal subset

## Facets
- artifact type: library
- maturity: active
- function: gpu-computing
- domain: gpu-computing, machine-learning, developer-tools
- platform: rust, windows
- tags: cuda, cuda-bindings, nvidia, cublas, cudnn, nccl, nvrtc, curand, cusparse, cufft, cupti, ffi-bindings, safe-rust, gpgpu, multi-gpu, dynamic-loading, gpu, linux

## Member repositories
- chelsea0x3b/cudarc (main) score 98

## Provenance
- Observed fields: from GitHub, fetched 2026-08-28T04:04:00.476344+00:00.
- Health v2: computed from the inputs above; adoption is never an input.
- Inferred fields (summary, facets, guidance): AI-extracted, prompt v1, taxonomy v1, on 2026-08-30T06:17:55.891717+00:00, confidence not recorded.
  - readme: https://github.com/chelsea0x3b/cudarc (fetched 2026-08-28T04:04:00.476344+00:00, sha 8887c672fcce)
  - registry_crates: https://crates.io/api/v1/crates/cudarc (fetched 2026-08-29T12:25:28.299938+00:00, sha c309aa2375ee)
- Data as of 2026-08-30T08:39:29.467469+00:00.