Author Topic: AMD can now run Nvidia CUDA on Windows or Linux thanks to ZLUDA  (Read 37 times)

Offline javajolt

  • Administrator
  • Hero Member
  • *****
  • Posts: 35044
  • Gender: Male
  • I Do Windows
AMD can now run Nvidia CUDA on Windows or Linux thanks to ZLUDA
« on: February 12, 2024, 09:24:12 PM »

It is no secret that AMD has had a hard time competing with Nvidia in comparison to how it does against Intel. The struggle for AMD is two-fold, while the company does have decent hardware, the software side is a bit lacking, especially in the field of productivity and workstations.

In recent years though, AMD has been making some effort to catch up with the behemoth that is CUDA (Compute Unified Device Architecture) with its own ROCm (Radeon Open Compute Platform). Under ROCm, AMD introduced HIP (Heterogeneous-compute Interface for Portability) which allows developers to translate CUDA source code to run on AMD hardware with the help of HIPIFY tools. Performance though is not quite on par with CUDA.

That is not the case with Project ZLUDA which, incredibly, allows AMD hardware to run unmodified CUDA applications, and the performance too is fairly okay for the most part. The latest release of ZLUDA, version 3, adds AMD support to the compiler.

The changelog headlined[color] "Nobody expects the Red Team" reads:

    Remove Intel GPU support from the compiler

    Add AMD GPU support to the compiler

    Remove Intel GPU host code

    Add AMD GPU host code

    More device instructions. From 40 to 68

    More host functions. From 48 to 184

    Add proof of concept implementation of OptiX framework

    Add minimal support of cuDNN, cuBLAS, cuSPARSE, cuFFT, NCCL, NVML

    Improve ZLUDA launcher for Windows

The developer notes the performance of the Radeon RX 6800 XT on OpenCL vs ZLUDA using Geekbench 5.5.1, and overall it's a close fight where the latter comes out on top on more occasions:

Phoronix tested the performance of ZLUDA to see how it fares against CUDA and also AMD's own HIP. We have only included Blender Classroom and BMW results as it is a fairly popular rendering application where GeForce seems to walk all over Radeon, especially with Optix:

As you can see in the images above, the ZLUDA seems to have put on a very good show in both Classroom as well as BMW scenes. It also outperforms AMD's own HIP in both instances. CUDA though, even without Optix, is still ahead, though the ZLUDA is impressive nonetheless.