Nvidia writes its CUDA translation ban into the installer

Nvidia's CUDA documentation now ships with a clause that bans using translation layers to run CUDA software on non-Nvidia hardware. According to Tom's Hardware, the clause has been in Nvidia's online EULA since 2021, but it appears for the first time in the documentation that installs with the toolkit, starting with CUDA 11.6. The restriction itself is not new; what changed is that you now accept it at install time, with no need to read the fine print online.

The context is straightforward. Because the CUDA-plus-Nvidia-hardware combination has proven efficient, a large share of GPU software depends on it, and as more competitive hardware hits the market, more people want to run those programs on other chips. There are two ways to do that: recompile the code, which requires access to the source, or run the already-compiled binaries through a translation layer. The clause targets the second option.

What the clause says

The installed EULA text reads: "You may not reverse engineer, decompile or disassemble any portion of the output generated using SDK elements for the purpose of translating such output artifacts to target a non-NVIDIA platform." In practice, that means you can still compile CUDA programs, but taking the compiled output and running it on another vendor's chip through a compatibility layer is off the table. The clause was absent from the EULA documentation installed with CUDA 11.5 and earlier releases, and present from 11.6 onward.

Who it looks aimed at

Per Tom's Hardware, the timing lines up with two sets of efforts. The first is ZLUDA, a translation layer for running CUDA binaries on non-Nvidia GPUs, which both AMD and Intel had taken part in before stepping back from further development, leaving the project floundering. The second is a group of Chinese GPU makers claiming CUDA compatibility: Denglin Technology advertises processors with a "computing architecture compatible with programming models like CUDA/OpenCL," and Moore Threads ships MUSIFY, a tool it designs to let CUDA code work on its GPUs. Whether MUSIFY is a full translation layer or partly a porting tool is not clear, and Nvidia has not commented on what the clause targets.

What still works

Recompiling CUDA programs for another platform remains perfectly legal, and it is the supported path. Both AMD, with ROCm, and Intel have porting tools that help move CUDA code to their respective stacks. And native code tends to run better than binaries pushed through a translation layer, which is exactly why the translation shortcut exists in the first place: it lets you run already-compiled binaries without the software vendor writing anything new for your chip. The trade-off is that it depends on a third party maintaining the layer, and it now runs up against the license.

Why it matters

If you are already running CUDA software on non-Nvidia hardware, the practical question is how you're doing it. Through a translation layer, that path is now explicitly out of bounds under the license; through ported, natively compiled code, nothing changes. For a team evaluating AMD, Intel or another GPGPU, the clause is a concrete checklist item: verify that the specific software in your workload is ported to the target platform rather than dependent on a compatibility layer. Nvidia says nothing public about enforcement, and this is a licensing term rather than a technical lock, so nothing stops the hardware competition itself. But writing the ban into the installer signals that Nvidia sees translation layers as a threat worth closing, and as rivals keep improving native support, the path that stays clean is the ported one.