Reportedly, AMD has provided financial backing to a project developed by a single individual known as ZLUDA. Initially designed as a drop-in CUDA implementation for Intel OneAPI, this project enabled CUDA applications to run directly on Intel hardware.

In 2022, AMD contracted Andrzej Janik to adapt ZLUDA for use on AMD GPUs with HIP/ROCm. Before this, Intel had considered developing ZLUDA but decided against it and did not provide funding for the project.

Andrzej Janik dedicated the last two years to adapting ZLUDA for Radeon GPUs, and the results are promising: numerous CUDA software can now run on HIP/ROCm without any modifications. Simply run the binaries as usual while ensuring the ZLUDA library replacements for CUDA are loaded.

V-Ray’s CUDA rendering capability for utilization on Radeon GPUs via ZLUDA.

However, AMD unexpectedly decided this year to halt funding for the project and not release it as a software product. Fortunately, there was a clause in the contract allowing Janik to open-source the work if the contract ended, which may happen in this case.

https://www.phoronix.com/review/radeon-cuda-zluda

As per Phoronix’s report, the project has resumed with the goal of enabling CUDA applications to operate on AMD hardware seamlessly, without the need for translation or code adjustments. Although not all applications run natively yet, such as NVIDIA Optix, developers can now execute binaries on Radeon GPUs without modifications.

The only prerequisite is to incorporate the ZLUDA library, which replaces CUDA.

In essence, this signifies that developers can now utilize CUDA support for applications not inherently tailored for AMD hardware. This enables the utilization of CUDA as a rendering API for software like Blender 4.0 or V-Ray. Particularly for Blender, which already features its Radeon HIP compute renderer, this implies that Radeon GPUs can achieve superior performance via the ZLUDA library compared to using HIP.

ZLUDA enables Radeon GPUs to execute native CUDA code within Blender 4.0 at a faster pace than AMD’s Radeon HIP code which I personally find highly amusing as it goes to show just how AMD’s software developers are seemingly incompetent as the company continues to invest heavily into countries such as India.

Additionally, ZLUDA facilitates faster execution of CUDA code compared to OpenCL code on AMD GPUs. However, it’s worth noting that certain applications may experience reduced performance when utilizing ZLUDA.

The recently open-sourced ZLUDA library has already shown promising outcomes. Although not all applications are currently supported, the project can now progress with contributions from additional developers.

The rationale behind AMD’s decision to stop funding the ZLUDA project remains unclear. However, as highlighted by Phoronix, the increased software support for AMD’s ROCm compared to two years ago may have influenced this decision.

Andrzej Janik, the sole developer behind this initiative, reportedly intends to persist in his efforts. Among the potential avenues to explore is the integration of NVIDIA DLSS functionality via ZLUDA for Radeon GPUs.

https://github.com/vosen/ZLUDA

The project’s official page has been revised to underscore its emphasis on AMD GPUs. Additionally, the developer is sharing internal benchmark findings, comparing them with OpenCL implementations.