Skip to main content

Rust CUDA August 2025 project update

· 4 min read
Christian Legnitto
Christian Legnitto
Rust GPU and Rust CUDA maintainer

Rust CUDA enables you to write and run CUDA kernels in Rust, executing directly on NVIDIA GPUs using NVVM IR.

Work continues at a rapid pace with significant improvements landing regularly. Here's what's new since our last update.

To follow along or get involved, check out the rust-cuda repo on GitHub.

Chimera demo blog post​

We published a blog post about our demo showcasing a single shared Rust codebase that runs on every major GPU platform. The demo uses Rust CUDA for CUDA support.

The post reached #1 on Hacker News and was popular on Reddit.

Rust toolchain updated​

Rust CUDA includes a compiler backend that compiles regular Rust code into NVVM IR. Because of this deep integration with compiler internals, Rust CUDA must use a very specific version of the Rust compiler. Rust CUDA now supports nightly-2025-06-23.

This aligns Rust CUDA with the Rust GPU project, which uses the same toolchain version. Having both projects on the same Rust version enabled the aforementioned demo to work with fewer hacks.

Migration to glam​

Maintainers @jorge-ortega and @LegNeato migrated from the vek math library to glam in PR #180. Glam is used by the Rust GPU project and this consistency enables easier code reuse.

While vek is still re-exported at cuda_std::vek, it is deprecated and will be removed in the future.

i128 support​

@LegNeato implemented emulation for i128 operations that aren't natively supported by the version of LLVM NVIDIA's tools are based on.

With this support, Rust CUDA's compiler backend can now correctly compile the sha2 crate from crates.io. We've added an example demonstrating the same sha2 crate used on both CPU and GPU.

Using unmodified crates from crates.io on the GPU is one of the unique benefits of using Rust for GPU programming.

Target feature support​

PR #239 added support for CUDA compute capability target features. Developers can now use #[target_feature(enable = "compute_75")] to conditionally compile code for specific GPU architectures, enabling better optimization and feature detection at compile time.

For more details, check out the documentation.

Added compiletests​

Previously we only verified that the project built in CI. GitHub Actions runners do not have NVIDIA GPUs, so we could not run tests to confirm correct behavior. This made changes risky because regressions could slip through unnoticed.

@LegNeato ported the compiletest infrastructure from Rust GPU to work with Rust CUDA. Compile tests let us confirm that the compiler backend behaves correctly and generates the expected code. While not full runtime testing, this change significantly improves reliability and makes regressions easier to catch.

Multi-architecture Docker images​

Rust CUDA uses a version of NVVM based on LLVM 7.1 and getting it set up manually can be tedious and error-prone. Rust CUDA's docker images aim to solve the setup issue. @LegNeato updated our Docker infrastructure to add support for ARM64.

Call for contributors​

We need your help to shape the future of CUDA programming in Rust. Whether you're a maintainer, contributor, or user, there's an opportunity to get involved. We're especially interested in adding maintainers to make the project sustainable.

Be aware that the process may be a bit bumpy as we are still getting the project in order.

If you'd prefer to focus on non-proprietary and multi-vendor platforms, check out our related Rust GPU project. It is similar to Rust CUDA but targets SPIR-V for Vulkan GPUs.