British company Spectral Compute has introduced the SCALE toolkit for GPGPU computing, which allows code for the Nvidia CUDA environment to run on AMD GPUs. The CUDA (Compute Unified Device Architecture) software platform is a closed ecosystem, although there are attempts to adapt it for computing on GPUs from other manufacturers. There is already an open-source project ZLUDA. Now SCALE has appeared.
The CEO of Spectral Compute believes that developers should be able to write universal code for any hardware platform. This has been the case for CPUs for many years, and why not for GPUs as well? SCALE will be the bridge that will open the way to such a universal software environment. It is a set of GPGPU tools, similar to NVIDIA CUDA, for running binary code on a GPU. SCALE does not translate code, but uses its own chain of compatible tools, which makes it adaptable to different platforms, including AMD GPUs. SCALE allows developers to work with a single version of the code and does not require the use of other languages, since it is source-compatible with CUDA. The authors say that they have already tested SCALE in Blender, Llama-cpp, XGboost, FAISS, GOMC, STDGPU, Hashcat and Nvidia Thrust on AMD RDNA 3 and RDNA 2 processors.
At the same time, the SCALE toolkit itself has a closed source code and is distributed under a free software license. It is also unlikely that Nvidia will like the appearance of such an application. The green corporation has previously included a ban on the use of ZLUDA and similar third-party solutions in its license agreement.
Source:
Wccftech
