It is common practice to migrate codebases from one project or programming language to another, as this allows developers to access a wider range of features and simplifies maintenance. In this case, however, there is an interesting development as Team Red has apparently accelerated its efforts to port the CUDA platform to run on AMD’s ROCm. In doing so, they have contacted a developer who has already been involved in a similar implementation.
AMD’s ROCm magic now supports CUDA APIs in a “workaround” method that involves the use of ZLUDA, an open source porting project originally developed to enable support for NVIDIA CUDA on Intel GPUs. ZLUDA faced technical and practical challenges that ultimately halted its development, especially when it came to interest from companies like Intel at the time. The developer behind it, Andrzej Janik, was contacted by AMD for an experiment, the results of which are now finally visible.
After extensive testing by Phoronix, ZLUDA was found to work almost perfectly with AMD’s Radeon graphics cards in conjunction with ROCm and NVIDIA’s CUDA libraries. It was amazing that no changes to the existing code were required. Although project development had stalled due to AMD’s apparent withdrawal, the work was kindly released as open source after the deadline, giving Phoronix the opportunity to conduct extensive testing.
Code porting opened up new possibilities, and Phoronix was able to successfully run Blender 4.0 with CUDA libraries. Surprisingly, the test results showed that NVIDIA and AMD were on par in terms of rendering performance. In addition, developer Andrzej Janik himself conducted benchmark tests on the performance of Radeon graphics cards with CUDA support, and the results obtained were extremely interesting.
There is a possibility that ZLUDA could develop enormous potential in the future. The platform combines the ROCm and CUDA stacks rather than separating them, allowing developers to utilize the specific capabilities of both platforms. Although native CUDA support on Radeon GPUs is unlikely, such resources could be very useful in AI development. The developer is optimistic about the future of ZLUDA and is testing NVIDIA’s upscaling capabilities with ZLUDA on RDNA GPUs.
Source: Phoronix
9 Antworten
Kommentar
Lade neue Kommentare
Urgestein
Veteran
Urgestein
Mitglied
Veteran
Urgestein
Veteran
Mitglied
Alle Kommentare lesen unter igor´sLAB Community →