NVIDIA Halts The Use of CUDA On Other Platforms, Lists New Warning in the EULA

wccftech.com:

NVIDIA has banned the use of CUDA libraries on other platforms like AMD & Intel, as the firm adds a new warning with CUDA's EULA.

NVIDIA Targets ZLUDA & Other CUDA-Dependant Solutions With Their Revised Policy, Ultimately Hindering Code Porting

While NVIDIA hasn't made any official statements to back this claim, Kernel/hypervisor Engineer, Longhorn discovered the warning and resorted to X to disclose the change. This step came after the CUDA platform witnessed an increasing adoption from third-party developers and companies who utilized NVIDIA's software powerhouse to upgrade their hardware capabilities.

Related Story HBM Manufacturers Witnesses Low Yield Rates, Making It Difficult to Pass NVIDIA’s Qualification Tests

The use of translation layers for running CUDA on other platforms was banned in 2021 when NVIDIA initially listed the EULA agreement. Still, the warning was explicitly present in it, making us wonder why Team Green decided to "suddenly" revise the EULA.

Well, if you are unaware of how other platforms are leveraging CUDA, a prime and recent example of it is the use of ZLUDA, which is an open-source library that effectively ports NVIDIA CUDA apps over to AMD's ROCm that does not require code adaption.

Using translation layers, ZLUDA's creator implemented CUDA libraries on the ROCm almost perfectly, which was astonishing for individuals but alarming for tech giants like NVIDIA simultaneously. Similarly, Chinese firms like Moore's Threads and many more have utilized some part of CUDA for development at their end, which isn't likable for NVIDIA.

Now, while code porting and the use of translation layers do hinder the fact that CUDA was solely developed for NVIDIA's own GPU solutions, and it does steal the "exclusivity" to a certain point, let's not forget that code porting implementations have the potential to expand the boundaries of computing, particularly in the domain of AI, since individuals can ultimately create a hybrid model of hardware and software