AMD Quietly Funded A Drop-In CUDA Implementation Built On ROCm: It's Now Open-Source

[dragon321]

is why CUDA is the de facto GPU computing standard. NVIDIA cares.

More like developers cares to optimize their code for market leader.

Topic: Nice work. However release notes explains that this projects was released only because AMD isn't interested in further development and project is de facto abandoned for now. Not a good sign in the beginning.

[avis]

A discussion on Hacker News among actual CUDA, OpenCL and ROCm programmers and users.

[bearoso]

AMD might no longer fund this, but I'm guessing a ton of other companies are looking into it now.

[avis]

"Złuda roughly means "delusion" / "mirage" / "illusion" in Polish, given the author is called Andrzej Janik this may be a pun " - source.

[Nille]

Impressive and its nice that AMD Card now can run CUDA Code. But one Information in that Article is disturbing to me. Its sound like that ROCm 6 is not compatible with ROCm 5. Is that correct?

[qlum]

I personally doubt it's because this would hinder ROCm support. If so, it would be pretty bone-headed.
As effectively this could take away control of CUDA from Nvidia, if most software works with ZLUDA, any breaking changes to CUDA could see more limited adoption. If ZLUDA is actually being adopted.

So my best guess would be legal issues or fundamental issues, where important parts of CUDA cannot be supported in a performant manner.

[DumbFsck]

I wonder if AMD abandoned it because Nvidia consistently performed better.
[...]

The reason I don't think this makes sense is that in some benchmarks ZLUDA did better than native ROCm / openCL / HID.

I think it doesn't matter that your solution is (with ZLUDA) always slower than your competition, if your solution is always slower than the competition in general (regardless of the code path).

My guess is they dropped this not because of how it looks for AMD nor for any fear of legal disputes, my guess is they think Andrzej is going to generate more value for them in another project, who knows, maybe even directly improving and optimising ROCm.

[stormcrow]

"The developer that coded this, originally coded the software to run on Intel GPU's, while he was an employee of Intel. So unless Nvidia fell asleep at the wheel, Nvidia didn't see any legal issues with the software. AMD certainly knew this" - source.

Oracle v. Google was decided in 2021. That case effectively makes APIs copyrightable in the US. While Intel began the project, that SCOTUS case alters the copyright field in the US. In order to use APIs, you either have to comply with the license, or be able to successfully argue fair use in a trial, an expensive and lengthy process without guarantees of success. People shouldn't make assumptions what AMD knows and doesn't know, or what Nvidia sees as a problem or doesn't. It's up to a property owner when they wish to file suit against alleged infringers. Just because they've been silent doesn't mean they will remain silent - nor do we know what's going on behind the scenes when no one talks. Successful estoppel defenses are rare. You can't assume a property holder is ok because "surely they know". You have to actively engage with the property holder in good faith, then and only then if they agree to allow something then renege on the agreement does one have a strong case for estoppel.

You should never take a tech reporter's assumptions about the law at face value, especially when they're making assumptions about who knew what and when. Reporters are nearly always inaccurate when it comes to legal matters. (especially in the tech press!)

[Developer12]

Not bad, why would AMD stop that project, especially considering the results?

The reason why ZLUDA is faster than HIP is most likely more optimizations in the CUDA path.

Because otherwise AMD would be building their empire on land that belongs to nvidia. nvidia still has outsize control over CUDA and could render CUDA applications incompatible with ZLUDA with a single version bump. It's kind of like what the kernel devs keep trying to do to nvidia wrt their driver module.

Much better for AMD and intel and so on to build their own ecosystem and try to get people to switch over. That's what AMD has done, producing HIP libraries that seek to be a drop-in replacement for nvidia's C++ cuda libs. intel is pursuing similar initiatives.

[DumbFsck]

Oracle v. Google was decided in 2021. That case effectively makes APIs copyrightable in the US. While Intel began the project, that SCOTUS case alters the copyright field in the US. In order to use APIs, you either have to comply with the license, or be able to successfully argue fair use in a trial, an expensive and lengthy process without guarantees of success.

I'm a layman in both law and programming, so take this for what you will.

Something I don't understand is that iirc the decision said something about copying APIs for increased compatibility, I remember this was mentioned when those guys made some messaging software compatible with apples imessage APIs and whatnot. But at the same time, I'm sure I remember something about copying to offer a competing solution. Which makes it very different for me to understand.


Also, how is this different than, let's say, DXVK? In principle, then, is Microsoft able to sue and get DXVK down since it copies and translates their API to something else?

I think this CUDA v. ZLUDA is more similar to DXVK than to apple's messages solution or Google's use of Java. (reminder that I'm a layman).

Either way I'm sure the only way to know all of this for sure would be to have it tested in court.

You can't assume a property holder is ok because "surely they know".

This is an absolute truth. See "portal 64",years of development in the open. Either Nintendo (because of how the game was developed) or steam (because of trademarks and assets) or both got it to stop on a dime, right when it was getting close to launch. They knew before then, most likely for years.