The compatibility issue with Nvidia CUDA11.6 and higher versions of EULA has attracted widespread industry attention.
On the evening of March 5th, Moore Thread released a statement stating that recently, the company paid attention to reports on Nvidia CUDA compatibility on the internet, and Moore Thread MUSA/MUSIFY was not affected. The report states that a clause in the EULA (End User License Agreement) for Nvidia CUDA11.6 and higher states that you cannot reverse engineer, decompile, or disassemble any results generated using this SDK and translate them on non Nvidia platforms.
CUDA is a programming framework based on C language launched by Nvidia, which allows developers to write and run general-purpose programs on GPUs (graphics processing chips). Nvidia GPUs are particularly adept at performing highly parallelized computing tasks, mainly relying on CUDA technology and a wide range of applications and developer communities built on CUDA.
An industry insider stated that Nvidia's statement prohibits third parties from using hack methods, allowing native CUDA programs to run directly on non Nvidia GPUs without source code modification and through runtime translation/translation. This mode poses a risk of infringement because it directly utilizes the interfaces, running intermediates, and outputs of the CUDA software stack.
Some analysts also suggest that the restriction seems to be aimed at preventing Intel and AMD from participating in recent plans such as ZLUDA, and perhaps more importantly, some GPU manufacturers in other countries from using CUDA code with translation layers.
According to the official website of Mole Thread, Mole Thread focuses on the research and design of fully functional GPU chips and related products, supporting various combinations of workloads such as 3D graphics rendering, intelligent computing acceleration, ultra high definition video encoding and decoding, physical simulation, and scientific computing, balancing computing power and efficiency. It can provide strong computing acceleration capabilities for China's technology ecosystem partners and widely empower multiple fields of the digital economy.
Actually, since 2021, Nvidia has listed online license terms prohibiting the use of translation layers to run CUDA based software on other hardware platforms, but previous warnings were not included in the documentation placed on the host system during the installation process. This description has recently been added to the EULA (End User License Agreement) included when installing CUDA 11.6 and higher.
"We hereby clarify that MUSA/MUSIFY is not related to the terms of NVIDIA EULA (End User License Agreement) and developers can use it with confidence." Moore Thread stated that MUSA is a fully functional GPU advanced computing unified system architecture independently developed by Moore Thread, with all intellectual property rights and integrated software and hardware, and has no dependency on CUDA.
It is reported that MUSIFY is a development tool provided by Moore's Thread for MUSA developers, which facilitates application porting and development on the MUSA computing platform. It allows developers to convert their C++source code into MUSA C++source code, and then compile it into binary code based on the MUSA instruction set through the MUSA compiler MCC. Finally, it runs on Moore's Thread's fully functional GPU.
Moore Thread expressed gratitude to the developers for their support. The company will fully leverage the advantages of its intelligent computing products, from chips to graphics cards to clusters, to continue polishing MUSA and providing high-performance computing support for large model training and inference.