NVIDIA CUDA can be used on Intel accelerators, but so far only in Geekbench

NVIDIA CUDA can be used on Intel accelerators, but so far only in Geekbench


Using GPUs for tasks beyond simple 3D rendering is the industry that has brought NVIDIA billions in the data center (and now mining) sector. Its proprietary CUDA platform and API have been exclusive to the company’s graphics cards from the start. But now the technology is available on Intel accelerators as well.

Of course, there have been tools in the past for porting CUDA applications to widely supported languages ‚Äč‚Äčlike OpenCL. However, even semi-automated tools like HIPCL required developer intervention. But the new ZLUDA library promises to run CUDA applications directly on Intel GPUs without any changes. ZLUDA uses Intel oneAPI Level Zero to translate or emulate CUDA commands.

The ZLUDA developers describe their creation as a direct replacement for CUDA on systems with Intel GPUs used in Skylake and later processors. The latest version supports both Windows and Linux with 9th generation graphics or newer. According to the developers, ZLUDA is able to achieve almost the same performance as if the code were directly compiled. However, there are still big limitations: the supported functionality is significantly limited. In fact, the library currently only really works on Geekbench, and potential users are warned not to rely on it for critical software development.

Using ZLUDA on Windows seems straightforward enough. To do this, download the latest build from the ZLUDA GitHub page. The downloaded archive includes a “wrapper” for running any CUDA-enabled application and the necessary library. Then you just start the shell from the command line with the application as an argument (zluda_with – geekbench5 –compute CUDA). However, Hot Hardware journalists gave this method an error.

On Linux, the author of ZLUDA was able to run benchmarks for the Core i5-8700K, scoring 6333 points with CUDA using integrated UHD graphics 630 versus 6482 points when executed through OpenCL. This is a slight overall decrease in performance, but there are some interesting nuances. Some Geekbench benchmarks were significantly slower, but for example the Stereo Matching benchmark was about 50% faster with ZLUDA than with OpenCL.

Comparison of test results in Geekbench via ZLUDA and OpenCL

Comparison of test results in Geekbench via ZLUDA and OpenCL

In other words, the creators of ZLUDA still have a long way to go – the library only went through the release of the second major build. However, this is an interesting project. Intel’s DG1 graphics cards have already been released for OEMs, and it seems that faster DG2s are just around the corner. While integrated graphics serve as an interesting proof of concept, there is no doubt that Intel discrete graphics are much more interesting in the future for ZLUDA.

If you notice an error, select it with the mouse and press CTRL + ENTER.

Leave a Comment