SCALE enables native compilation of Nvidia CUDA code for AMD GPUs

Nvidia CUDA is a computing architecture that runs exclusively on Nvidia graphics cards. Years ago, it used the GPU’s ability to process large amounts of data with the same operations at once to accelerate some scientific and other computational tasks. Over the years, it has become an industry standard that the competition is unable to break. AMD was not very successful with its first attempts, there was also an effort to expand the general OpenCL standard, but recently the ROCm alternative has managed to get more attention. But there is still the idea of ​​running codes written in the CUDA interface even on AMD graphics cards. There was, for example, the ZLUDA project, which was financially supported by both AMD and Intel. The latter used translation layers, which is something that Nvidia prohibits in its terms and conditions.

But now, after 7 years of development, a new project has appeared from Spectral Compute SCALE, who should catch this by the right end. It will not rely on any translation layers to translate the compiled code for AMD cards, but can directly compile CUDA code for AMD graphics cards. The company thus hopes that CUDA will become a similarly more open platform, where it does not matter so much about the manufacturer of the chip on which the given code will run. The head of the company likened it to the situation with processors, where x86 code runs on both Intel and AMD. SCALE basically replaces nvcc and looks like the Nvidia CUDA Toolkit with support for all possible compiler calls that are supported for Nvidia cards. The result is similar, the only difference is that the resulting code will run on AMD cards.

It has already been tested on applications such as Blender, Llama-cpp, XGboost, FAISS, GOMC, STDGPU, Hashcat, or NVIDIA Thrust and works on cards with RDNA 2 and RDNA 3 architectures. So these are common consumer cards, CDNA computing architectures are not yet supported. And here we know very well that the computing area is a market where Nvidia really does not want competition, because it is precisely things related to computing cards (and artificial intelligence) where it has a dominant position and where its revenues have increased by an order of magnitude in recent times.