Developer Speedstu published CUDA-for-AMD-Windows on GitHub on Sept. 13, bundling ZLUDA and AMD’s Windows HIP SDK into a PowerShell installer designed to run CUDA software written for NVIDIA GPUs on AMD Radeon graphics cards.
The project is not a new translation layer by itself. ZLUDA is the underlying open-source component that handles CUDA calls and redirects them to AMD’s HIP runtime. Speedstu’s contribution is the surrounding workflow: version pinning, downloads, hash verification, environment setup, and diagnostic tooling packaged into a repeatable installation process.
Validation so far is limited to the RX 9060 XT
The documented test environment uses ZLUDA v6-preview.69, AMD HIP SDK 6.4, and LibTorch 2.3.0+cu118 on an AMD Radeon RX 9060 XT with the gfx1200 architecture. The README says this is the only GPU that has been validated so far. Other AMD cards are listed only as candidate devices, with no guarantee that the setup will work.
In that environment, ZLUDA’s built-in cuda_check tool reported five working components: nvcuda, cuBLAS, cuBLASLt, cuSPARSE, and cuFFT. The mappings described in the documentation route cuBLAS to rocBLAS, cuBLASLt to hipBLASLt, and cuSPARSE to rocSPARSE. These libraries cover operations such as matrix multiplication, sparse matrix handling, and Fourier transforms, all common in AI and scientific computing workloads.
Speedstu then tested the setup with an existing training program. The workload was a PPO reinforcement learning model with 2,216,347 parameters. According to the README, the model completed inference, learning updates, and optimizer calculations, finishing one run of 65,536 timesteps. The documentation says this was the original training workload that led the developer to build the project. Another public repository from the same developer uses LibTorch and PPO to train a Rocket League game AI.
Work to broaden hardware support has already started. On Sept. 14, another developer, l33tm4st3r, submitted a merge request adding support for the gfx1201 architecture. The card used in that test was an AMD Radeon AI PRO R9700, though the change has not been merged. The same gfx1201 architecture is also used by the RX 9070 series. AMD’s official Windows hardware support table says the RX 9070 series and RX 7000 series both support the HIP SDK, but whether those GPUs can run this workflow still depends on user test reports.
Performance figures come with caveats
A comparison test dated Sept. 13 used the same training setup to measure the official ZLUDA path against an earlier set of locally modified components. Each path ran two groups of five rounds. After removing the first warm-up round from each group, eight rounds remained for each side. The official path posted a median throughput of 13,278 steps per second, while the modified version delivered 12,876 steps per second, making the modified path 3.03% slower. The project therefore uses the official version by default.
The source code for that older modified component set is no longer available. The project notes say the developer recovered only compiled DLL files from an earlier development environment. Source code and debugging files for the cuBLASLt bridge layer were lost, so only hash values for those files have been retained and not published.
The README also lists older records ranging from 70,000 to 109,000 steps per second, but it explicitly says those numbers came from a heavily tuned and different configuration, so they should not be compared directly with the benchmark above. The same documentation adds that the developer later rewrote the training program to call AMD HIP directly, removing both LibTorch and ZLUDA, which produced much higher speed. The developer’s own training code no longer runs through ZLUDA.
Support limits and licensing constraints remain
The project’s limitations section makes clear that ZLUDA is not a full CUDA implementation, and the Windows HIP SDK exposes only part of ROCm. The stable HIP SDK on Windows does not include MIOpen, so software that depends on cuDNN may fail. cuDNN is NVIDIA’s deep learning acceleration library and is widely used by image models that rely on convolution-heavy workloads.
The same list says NCCL for multi-GPU communication, TensorRT for inference acceleration, and some custom CUDA extensions may also fail. The green "verify passing" badge on the project page does not mean CUDA programs have passed on AMD hardware. That GitHub Actions workflow runs on GitHub’s Windows cloud hosts and checks PowerShell syntax, JSON configuration formats, scan tools in an environment with no GPU, and hash values for downloaded ZLUDA archives.
The pinned versions are not the newest ones available. The project is locked to ZLUDA v6-preview.69, released on May 4, 2026. ZLUDA then shipped its stable v6 release on June 29 and reached v7-preview.10 on Aug. 26. AMD’s latest Windows HIP SDK is version 7.2.0, released on July 31, while the project validation used version 6.4. The README says newer versions may work, but they remain unverified until users report results.
Licensing is another unresolved issue. NVIDIA’s CUDA license terms still include a translation-related restriction that prohibits reverse engineering, decompiling, or disassembling programs generated by CUDA development tools for the purpose of making them work on non-NVIDIA platforms. The latest update to that license is dated Jan. 26, 2026. The project’s third-party notice also says NVIDIA components remain subject to NVIDIA’s licensing terms.
ZLUDA funding has weakened, and compatibility reports are still absent
ZLUDA’s own funding picture is unstable. The project had AMD backing starting in 2022, but that contract ended in early 2024. After that, the codebase was rolled back to a state from before AMD’s involvement and was later supported by an unnamed company. In an update post, developer Andrzej Janik said ZLUDA no longer has commercial funding and has returned to being a personal weekend project. He said his priorities have shifted toward work he personally finds interesting, including PhysX and Blender support.
Two days after CUDA-for-AMD-Windows went live, there were still no GPU compatibility reports on GitHub. For now, there is no public test data showing whether models beyond the RX 9060 XT can run this setup.
What is known at this stage
- The project aims to run some CUDA programs on AMD GPUs under Windows.
- The only validated GPU so far is the AMD Radeon RX 9060 XT.
- Five checked components passed: nvcuda, cuBLAS, cuBLASLt, cuSPARSE, and cuFFT.
- cuDNN is still unavailable, so software that depends on it may not run.
- RX 9070 series and RX 7000 series cards support the HIP SDK, but this workflow has not yet been publicly verified on those models.
FAQ
Can AMD GPUs run CUDA programs now?
Partially. Public documentation shows the project passed checks for five CUDA-related components on the RX 9060 XT and completed training for a model with 2,216,347 parameters, but software that needs cuDNN may still fail.
Which GPUs does CUDA-for-AMD-Windows support?
At this point, only the RX 9060 XT has been validated. A Sept. 14 submission added gfx1201 support based on testing with the AMD Radeon AI PRO R9700, but that change has not been merged.