TECH Signal 615 2 feeds carried it
Public ZLUDA stack enables CUDA applications to run on AMD GPUs in Windows
A reproducible Windows setup using ZLUDA and AMD HIP/ROCm allows CUDA-targeted applications to execute on AMD GPUs, validated on RX 9060 XT hardware.
Engineers can now test CUDA workloads on AMD GPUs without NVIDIA hardware, reducing vendor lock-in for compute applications. However, support is limited to specific hardware and may not cover all CUDA APIs or workloads. This provides a practical alternative for developers constrained by GPU availability or cost.
Written by elseif from the cluster below · every claim links back to a sourceThe three things worth knowing
The stack combines ZLUDA with AMD HIP/ROCm to translate CUDA calls for AMD GPUs on Windows.
Validation is confirmed only for AMD Radeon RX 9060 XT (gfx1200), with other GPUs untested or unsupported.
cuDNN is unavailable with the stable Windows HIP SDK, limiting compatibility for convolution-heavy workloads.
THE READ
What the cluster adds up to.
This development introduces a public, reproducible method to run CUDA-targeted applications on AMD GPUs under Windows. The stack leverages ZLUDA to intercept CUDA calls and translate them into AMD’s HIP/ROCm ecosystem, enabling compatibility with libraries like cuBLAS, cuBLASLt, cuSPARSE, and cuFFT. The setup is validated on a single GPU model (RX 9060 XT, gfx1200), with no guarantees for other hardware. Engineers must treat this as an experimental path rather than a drop-in replacement for NVIDIA GPUs.
Adopting this stack requires installing AMD’s HIP SDK and running a PowerShell-based installer that configures ZLUDA and LibTorch. The process is automated but involves downloading large dependencies (e.g., LibTorch at ~2.66 GB) and verifying SHA-256 hashes. The installer also generates runtime configuration files and runs diagnostic checks to confirm compatibility. While the stack avoids private or recovered DLLs, it relies on upstream components, meaning future AMD SDK updates could break compatibility until tested.
Performance benchmarks show the public upstream path is marginally faster (~3%) than a custom overlay for a PPO workload, but this is workload-dependent. The stack does not support cuDNN with the stable Windows HIP SDK, which may exclude convolution-heavy applications like certain deep learning models. Engineers should expect gaps in API coverage and potential instability, particularly for workloads outside the validated scope. The project’s documentation explicitly notes that not all CUDA programs or AI models will work.
The practical value of this stack lies in its ability to unblock CUDA workloads on AMD hardware, particularly for developers without access to NVIDIA GPUs. However, its narrow hardware validation and missing cuDNN support limit its immediate utility for production environments. Engineers should test their specific workloads thoroughly, as success depends on the CUDA APIs used and the underlying GPU architecture. The project’s open-source nature allows for community contributions to expand compatibility.
Written by elseif from the cluster below · checked for specifics the sources never containedTHE CLUSTER