Hacker News Reader

Comment by sandworm101 | Hacker News Reader

New idea: Break the cards.

A gamer doesn't need mathematical perfection. 99% of gamers wouldn't care if a card game rendered something slightly off, a pixel or two out of place on a 8k screen. A slight mathematical error wouldn't be noticed. But that same math error would destroy any mining efforts. So I propose that, rather than feeble attempts at software-based DRM, that Nvidea actually break the cards. Add in a tiny math error that gamers won't care about but that will render the cards useless for anything requiring exact calculation.

This would of course be very difficult to design.

View on HN

Modern engines render scenes in multiple passes, iterating over the same framebuffer until everything has been rendered. Engineering such a flaw in a way that wouldn't lead to dramatic cascading effect will indeed be very difficult to design.

Note that sometimes games do need absolutely exact results, or at the very least reproducible results. For instance sometimes you want to use something like `glDepthFunc(GL_LEQUAL)` which will only accept a fragment if its depth is equal to the depth buffer's value. A small fudging here would cause potentially large visual issues.

I'm not saying it's impossible but I expect that it would create a lot of headache as random, potentially unmaintained games would start glitching here and there.

Look at the Dolphin blog posts for loads of examples of where minor math errors add up to completely break a game.

The other big use and market nVidia wants to have is ML installations for data centers though. Those also want perfect math and are pretty big business for nVidia. Also no one is running an 8k monitor off a 3060 or particularly well off of a single card usually. Also I dread trying to track down a rendering bug and finding out it's an intentional defect in cards, effects are getting pretty intricate and with raytracing small errors will accumulate over the multiple bounces.

Intel managed this accidentally with the FDIV bug: for certain very specific values, floating point division gave the wrong answer.

Then there was the traditional rigged demo solution; one generation of cards behaved differently when the program running was named QUAKE.EXE. That is now almost standard, since part of the reason NVIDIA driver downloads are so large is a huge pack of compatibility tweaks for specific games.

I'm always amazed how consumers will demand less for their money.

I definitely wouldn’t buy something that had been intentionally crippled, on principle. Even if it were a very slight crippling.

Which would also make the cards useless for any GPGPU workloads.

Isn't that exactly what they want? To only allow the cards to be used for gaming, not compute tasks like mining.

The problem with that however is that games might be using the GPU for non-visual things such as physics or AI. Would those 'off by one' errors impact those tasks?

Yea. But in a good way.

That might actually suit NVIDIA's interests just fine - they could sell uncrippled GPGPU-capable cards at a premium whilst claiming that they're helping to protect the supply of cards for gamers. However I think it would be difficult to do this without also breaking rendering pipelines.

This would split the market for ML as well.

Approximate ML is a hot research topic.

Reply on news.ycombinator.com