Hacker News Reader

Someone with early access to 3060 hardware, but not access to the launch drivers managed to wrangle it into working on pre-3060 drivers and still found that mining performance was artificially reduced. It starts high then falls off after a minute or so when some heuristic decides that mining is happening.

https://videocardz.com/newz/zotac-geforce-rtx-3060-early-gpu...

There's speculation that the limiter might actually live in the cards VBIOS, which is much harder to modify than the driver as newer Nvidia cards require a signed VBIOS to boot.

I'm really curious about how they got the card working without the specific drivers for it.

I'm also curious if this "mining detection" is going to kick in "accidentally" on other compute loads. It would be really crappy if the CUDA performance of these cards is bad because they don't want miners using them.

If you watch the Russian source video he seems to somehow install Nividia's drivers selecting 3070.

So we’ll see mod chips for GPUs soon?

YES! I hope so if only to feel like I am back in the scene for just a moment. Miss those days. Maybe they can mod the VBIOS to show the crack group's logo on bootup just like how old VGA cards would show some company text before handing off to the PC.

> old VGA cards would show some company text before handing off to the PC

My Gigabyte motherboard is doing that. Ain’t too old, has AMD’s X570 chipset.

Well, that's the motherboard, not the videocard. Every motherboard does it.

Yep, that's through the BGRT, which is part of ACPI. I'm not actually sure if a GPU could export one to the BIOS, but I assume not.

The connection to the monitor is physically located on the video card. So the gpu could display whatever it wants without asking or coordinating with the motherboard.

Or the CMP line will actually fit the needs, and no need for games of whack a mole.

It basically depends on three things though. The CMP card supply needs to be sufficient for the miners' demand, they can't be more expensive, and the question of how much does being able to resell the cards to gamers play into the value proposition.

I actually avoid the used video card market now, the resale price is insanely high and odds are the card was driven as hard as possible for months at a time mining crypto before being swapped out for the latest and greatest.

Follow-up: Bryan Del Rizzo from Nvidia says the limiter involves the cards firmware: https://twitter.com/bdelrizzo/status/1362619264423747590

I don't speak Russian but the source video shows him force installing Nividia's drivers and selecting 3070.

Pure speculation but I'd imagine that Nividia's drivers can properly differentiate and already had the limitation in the current release.

Does this also happen during machine learning workloads?

Nvidia is effectively incentivizing the cracking of their driver and firmware security signing with millions of dollars. Because a person who cracks it can build a more profitable mining farm

It's a sort of accidental bug bounty and I wonder if Nvidia have really thought this through?

If you crack it, you haven't found a bug though. You managed to do a bit of circumvention.

I'd see locked firmware as a bug. Nvidia probably wouldn't agree.

Plenty of Nvidia Linux users will accept the "patch" regardless I think.

Indeed, I remember when I was doing some mining early in bitcoin's history. I had custom hacked BIOS on my AMD cards to be able to overclock beyond regular power limits. This flashing was possible only due to an exploit of the BIOS signature verification.

Given that I was doing this as only a rather casual miner with three GPUs, I have no doubt that this new restriction will be attacked ruthlessly.

Is it even in theory possible to produce silicon that does not perform well on crypto operations while allows 3D to run at full speed? Or do they simply rely on the same underlying operations?

wouldn't hashing be mostly int operations, while 3D graphics would be FP32? I dunno if they still do this, but nvidia used to artificially limit the rate for other float widths on geforce parts to segment the market.

There's plenty of integer ops in games these days. They do pretty arbitrary computation in compute shaders.

> I already do it for the video encoding limitations, why not this too?

That sounds pretty interesting. Mind elaborating on what changes you achieve with your modifications?

https://github.com/keylase/nvidia-patch

Basically just lets my Plex server transcode more streams at once without having a Quadro.

> my Plex server transcode more streams at once

And... why do you need your Plex server transcoding multiple streams? Just curious!

I've got many users on it and 3 or 4 simultaneous streams is not an uncommon occurrence. Nvenc is artificially limited to 2 on consumer cards.

Thought I was reading https://developer.nvidia.com/video-encode-and-decode-gpu-sup... wrong but turns out they raised the limit to 3 last year.

https://techgage.com/news/nvidia-bumps-nvenc-stream-limit-on...

Its not question if miners can/will crack it at some point (yes, they can). Question is how many miners will order the cards on day 0 without knowing if they can be used? The intention here is only to deter them enough that gamers have at least slightly better chance of getting these cards, and especially for launch-day orders.

I guess as long as mining shitcoins on a gpu is profitable we will have the gpu supply problem.

Maybe it's time to create an adversarial network that can figure out how to mine on the GPU without causing it to halve the hash rate, train that network on an NVIDIA GPU, and shove it to them.

That's not how adversarial networks work