Microsoft: "what do you mean, your PC?"

I read the article, and a few points stuck out to me:

1. This has been a restriction since 2021; now it’s documented in the files and not just the online EULA (ie consistent)
2. This is a protection to disallow other companies like Intel and AMD from profiting off of Nvidia’s work
3. Nothing is stopping anybody from porting the software to other hardware, eg

Recompiling existing CUDA programs remains perfectly legal. To simplify this, both AMD and Intel have tools to port CUDA programs to their ROCm (1) and OpenAPI platforms, respectively.

I’m all for piracy and personal freedoms, but it doesn’t seem to be what this is about. It’s about combating other companies profiting off Nvidia’s work. Companies should be able to fight back against other companies (or countries).

I mean it’s not like Nvidia is unreasonably suing open-source projects into oblivion or anything, or subpoenaing websites for user data; at least, not yet.

Agreed. I hope lawmakers step in and make EULAs like this unenforceable.

It probably is. In the EU APIs aren't copyrightable in the first place, doubly so if it's necessary for interoperability, in the US there's Google vs. Oracle which declared Google's use of Java APIs in Android fair use.

That's the neat thing about being in the American oligarchs class. If it's illegal just make it legal.

Feels like a fantastic base for an anti-trust case at the least.

Nvidia is dominating the AI chip market. If our laws were properly enforced, Nvidia should’ve been too afraid to abuse their market position like this.

I am extremely tempted to @ some guy who was shilling for nvidia and saying they weren't a monopoly.

Wine is done on clean room reverse engineering, it doesn't use any propetriary code as reference. If they had done so, Microsoft would have grounds to sue them.

This can't enforce anything on CUDA versions below 11.6; but any functionality introduced to CUDA after 11.6 needs to be clean room reverse engineered, so this will make ZLUDA development on those versions more difficult.

It might not be. But the mere potential of having to litigate for years will have a chilling effect.

https://www.eff.org/issues/coders/reverse-engineering-faq

Yes, the clause might be unenforceable on fair use grounds. So, if you feel like going through a couple years of risky litigation...

Funny how people aren't cheering on NVIDIA.

You can use graphics cards for more than just graphics, eg for AI. Nvidia is a leader in facilitating that.

They offer a software toolkit for developing programs (an SDK) that use their GPUs to best effect. People have begun making "translation layers" that allow such CUDA programs to run on non-nvidia hardware. (I have no idea how any of this works.) The license of that SDK now forbids reverse engineering its output to create these compatibility tools.

Unless I am very mistaken, Nvidia can't ban the use of "translation layers" or stop people making them, as such. This clause creates a barrier to creating them, though.

Some programs will probably remain CUDA specific, because of that clause. That means that Nvidia is a gatekeeper for these programs and can charge extra for access.

It's not about it being fast, it's about it only being available for NVidia GPUs. As long as software for things like machine learning uses CUDA, you need to buy an NVidia GPU to use it. A translation layer would let you use the same software on other companies' GPUs, which means people aren't forced to buy NVidia's GPUs anymore.

The EULA is associated with the CUDA software, not the NVIDIA hardware.