ZLUDA: A CUDA Compatibility Layer for Running NVIDIA CUDA Applications on AMD GPUs

Forum topic · 小凯 · 2026-03-06

Summary

ZLUDA is an open-source CUDA-on-AMD compatibility layer developed by Polish engineer Andrzej Janik (vosen) that translates NVIDIA CUDA calls into AMD ROCm/HIP, enabling many CUDA applications to run on AMD Radeon RX 5000-series and newer GPUs at near-native performance. Originally built for Intel GPUs, the project pivoted to AMD after undisclosed funding in 2024, which was paused in February 2024 as AMD shifted focus to ROCm v6. Development continues independently, with the team expanding to two full-time developers in Q2 2025. Recent milestones include fixing the ROCm 6.4 comgr ABI break, GitHub automated builds, and a jump in llm.c compatibility from 0 to 552. ZLUDA requires Rust 1.66.1+, CMake, Python 3, ROCm 5.7+ on Linux, or AMD Radeon Software on Windows.

Overview

ZLUDA is a CUDA-to-ROCm/HIP compatibility layer that allows NVIDIA CUDA applications to run on AMD GPUs at performance close to native. The project translates CUDA runtime and driver calls into their AMD ROCm/HIP equivalents.

Core Information

| Attribute | Detail | |---|---| | Author | Andrzej Janik (vosen), Polish developer | | GitHub | https://github.com/vosen/ZLUDA | | Underlying tech | ROCm / HIP | | Supported GPUs | AMD Radeon RX 5000 series and newer (desktop and integrated) | | Language | Rust (1.66.1+) | | Status | Alpha quality, under active development |

History

1. Originally developed as a CUDA compatibility layer for Intel GPUs. 2. In 2024, AMD quietly funded the project, which then pivoted to AMD GPU support. 3. In February 2024, AMD discontinued funding as it shifted focus to ROCm v6, but permitted the author to open-source the code. 4. In Q2 2025, the team expanded to two full-time developers and continued active development.

Usage

Linux:

LD_LIBRARY_PATH=ZLUDA_DIRECTORY APPLICATION ARGUMENTS

Windows:

ZLUDA_DIRECTORY\zluda.exe -- APPLICATION ARGUMENTS

Alternatively, copy nvcuda.dll and nvml.dll from the ZLUDA directory into the application's folder.

Confirmed Supported Applications

Geekbench
Blender 4.0 (reported 10–20% faster than native ROCm/HIP)
3DF Zephyr
Reality Capture
LAMMPS, NAMD
waifu2x
OpenFOAM
Arnold (proof of concept)

Performance

Up to 75% faster compared to OpenCL implementations.
Blender rendering reported 10–20% faster than native ROCm/HIP.
Design goal: near-native CUDA performance.

Technical Limitations

No support for 32-bit processes.
Partial support for OptiX and PTX assembly code.
Some vendor performance libraries are unsupported.
Limited support for llama.cpp and Arnold.

Latest Progress (Q2 2025)

Team expanded from one to two developers.
Resolved the ROCm 6.4 comgr ABI breakage.
Improved compiler correctness.
Added GitHub automated builds.
llm.c support rose from 0 to 552 (compilable/testable workloads).

Build Dependencies

Git
CMake
Python 3
Rust 1.66.1+
C++ compiler
ROCm 5.7+ (Linux) or AMD Radeon Software (Windows)
Ninja (recommended)

Clone and Build

git clone --recursive https://github.com/vosen/ZLUDA.git cd ZLUDA cargo xtask --release