Hacker News Reader

Comment by 542458 | Hacker News Reader

GPUs have been doing this forever. This is exactly the same as some features being enabled on quadros that don’t exist on RTX/GTX/GT cards. That’s (mostly) not silicon, that’s just firmware.

This isn’t covering any supply chain failures (despite what nvidia says, the number of available GPUs remains approximately fixed), it’s just more market segmentation to minimize the secondhand mining GPU market and capture more dollars.

I was also concerned about detection false positives, but as far as I can tell they’re detecting a very specific algorithm that only gets used for ETH mining - there have been no false positives that I can see. Some miners are even working around the limiter by introducing obfuscation operations, but that comes with a hashrate cost of course.

View on HN

> GPUs have been doing this forever though. This is exactly the same as some features being enabled on quadros that don’t exist on RTX/GTX/GT cards.

No, I think this is the first time a GPU has been heuristically trying to detect what you're doing and self-limiting in response. How is that equivalent to features just being literally off?

There's no special "mine-crypto" instruction in the CUDA ISA.

> there have been no false positives that I can see.

How can you see that? These aren't released yet and the GTX-3060 isn't really a popular card among serious gamers that would notice.

The 3060 has been out for a few months now, and it has the limiter. Re: your edit about noticing, I strongly disagree. Tech reviewers would LOVE to run the “limiter cripples Blender/Tensorflow/whatever” story, but that hasn’t happened. And in this market users are generally taking what they can get, so I’ve seen lots of people who would normally run 70/80/90s running 60s instead.

The existing limiter doesn't work, though. It was narrowly focused and consequently easy for miners to work around. That's why they're making this new version. There is no precise test for whether a given workload is "mining", though, so we can expect more false positives to show up as Nvidia tries to close off the myriad ways to avoid detection and miners respond by making their computation look more like traditional gaming workloads.

> No, I think this is the first time a GPU has been heuristically trying to detect what you're doing and self-limiting in response

nah, remember when NVIDIA decided that Titans were now prosumer cards and would get Quadro series drivers, and then performance tripled in CAD applications once the limiters were removed?

https://wccftech.com/nvidia-titan-xp-titan-x-385-12-driver-u...

The distinction between a Quadro and a GeForce has literally always been drivers and firmware. NVIDIA would have the card look at the specific API calls being made and if it decided you were a CAD application then it'd slow itself down on GeForce cards, while running normally on Quadro.

AMD does the same thing with Radeon Pro, and Frontier Edition cards, and so on.

You literally don't understand that this has been the norm for a decade now. Miners are just crying because they're finally being asked to purchase a business segmented product for their business venture.

The number of GPUs being produced may be constant but it’s lower than desired. The issue is Nvidia worked with Samsung to fab their chips this time around and Samsung had poor yields. Nvidia should have used TSMC like AMD and Apple but they tried to use Samsung and got burned on supply. They said they will use TSMC going forward but in the meantime there are shortages.

Reply on news.ycombinator.com