Hacker News Reader

Comment by serf | Hacker News Reader

>Funny how it turned out to just be a driver thing.

sort of makes one wonder how many whizz-bang hardware features get advertised that are little more than just software, but are oversold as unique physical engineering methods/techniques.

View on HN

I understand that pretty much the only difference between $3000+ "professional" GPUs that get certified for AutoCAD and other workstation systems are identical to $1000 "gaming" GPUs, just with a different on-card BIOS/firmware and e-fuses. I remember ages ago that people were able to simply re-flash their ATI Radeons into FirePro cards and that unlocked higher performance in some applications.

Same thing with Cisco switches for port-unlocks and Tesla cars (EAP and FSD are just software features, assuming you have HW2+, and rear heated seats are standard, just not activated unless you pay to enable it if you didn't get the cold-weather package when you ordered it).

sorta similar to cars that can get over the air updates and suddenly accelerate from 0 60 faster then before...

Yeah, little known fact. The Tesla Model 3 Performance and the Model 3 LR AWD have the exact same batteries and motors. There's literally nothing stopping Tesla from offering an even higher performance boost for LR AWD owners to grant the same 0-60 as the Performance model, but they don't because it would cannibalize Performance sales.

Download more ram, but real. What a world live in.

I wouldn't say nothing. A higher power profile will cause more wear, leading to more warranty claims. TANSTAAFL.

That's okay, warranties exist for a reason.

The warranty is likely to expire before the higher wear actually starts leading to reliability issues.

I'm also unsure how much it really makes a difference in an electric motor compared to a combustion engine. Unless you're going to the track (or significantly breaking speed limits), you're only putting extra wear on the motors for 2-4 seconds at a time, which is all it takes to get to your 40-70 mph target speed.

I present you with Flir 80×40 resolution E4, one setting away from full 320×240 E8

https://hackaday.com/2013/11/04/manufacturer-crippled-flir-e...

Oscilloscopes are another product notorious for shipping same hardware with different jumpers/variables in the firmware.

https://tomverbeure.github.io/2020/06/27/In-the-Lab-Tektroni...

somewhat famous DS1052E 50 to 100MHz hack https://www.youtube.com/watch?v=LnhXfVYWYXE

DS2072A 70 to 300MHz https://hackaday.io/project/8115-rigol-ds2072a-upgrade-to-ds...

MSO5000 70 to 350 MHz https://www.eevblog.com/forum/testgear/hacking-the-rigol-mso...

Keep in mind, Nvidia does have some genuinely good process mastery around handling encryption intended to keep users from utilizing hardware in ways they don't intend, and creating trusted computing platforms where safety critical systems are concerned.

https://docs.nvidia.com/drive/drive_os_5.1.6.1L/nvvib_docs/i...

They've moved from keeping that sort of thing in EEPROM's to burnable fuses from what I understand, and I'm pretty sure what I'm aware of is pretty far behind state of the art.

If they've managed to set up the key management as well as they have and keep things hush-hushed enough to keep the nouveau folks obstructed, this seems either like an uncharacteristically careless mistake, or some seriously well executed malicious compliance from somebody.

Either way. I still find it irritating to the extreme that this type of thing only seems to happen to the detriment of users. Nvidia wins either way. Miners or gamers will buy out their cards.

It is my understanding Nvidia allowed cryptominers to pull up semis and load them up with GPUs in exchange for about a billion dollars.

My guess is this was accidentally on purpose (see also the recent shareholder lawsuit; Nvidia did win that one btw).

> It is my understanding Nvidia allowed cryptominers to pull up semis and load them up with GPUs in exchange for about a billion dollars.

Source?

Never heard such a thing. Partners (MSI, ASUS, etc): absolutely. NVIDIA, I've never heard any evidence they sold directly to miners.

> My guess is this was accidentally on purpose (see also the recent shareholder lawsuit; Nvidia did win that one btw).

I think it's incredibly likely that we see this re-introduced on 3080 Ti or 3080 Super or whatever they call it which is about to be released, as well as with future generations of cards, which strongly argues against this being an "accidentally-on-purpose" thing.

People love to make up conspiracies about NVIDIA but all it takes is someone building a beta driver from a feature branch based on an old build that didn't have the mining brake yet, it's quite probable that this is just an accident. But it's NVIDIA and they're the only company people love to hate more than Apple.

People hate NVIDIA for releasing drivers that broke the mining brake, and they'll hate NVIDIA for reinstating the mining brake in the future. People get mad no matter what NVIDIA does.

I'm assuming they're implying that this "theft" was something else: https://arstechnica.com/gadgets/2020/12/340000-of-nvidia-rtx...

Which is still partners. And that's like, a single container getting lost. $340k is a far cry from billions as well.

I'm assuming he's referring to an RBS article which was mischaracterized by Barrons and then widely broadcast in the tech media... the RBS article looked at hashrate growth since launch and attempted to estimate the number of 30-series GPUs sold (the number they came up with was about 7% of gaming revenue). Barron's then took that number and trumpeted it as NVIDIA selling cards directly to miners when that's not what the original article said at all, they were just estimating how many cards ended up mining from any source, and it was sub-10% regardless.

But of course nobody actually went back and looked at the original article, and because it's NVIDIA everyone is completely willing to give full faith to any claims that make them look bad. Lie goes around the world before the truth has its pants on, etc.

But if there's an article I'm willing to read it.

I was indeed referring to the RBS article, which as you mention was bunk. I read it in passing several months ago. And it was $175 million they estimated, I must've crossed wires with the recent $1B shareholder suit.

Thank you for making me do a little leg work. Perhaps I should have done that in the first place.

Thanks for being open-minded!

Finally dug it up, source methodology from the original RBS article:

https://twitter.com/dylan522p/status/1332502890104188929/pho...

It's too easy to hate a company like nvidia. Even if you make things up, they're probably guilty of them.

I'll start to like nvidia when they provide anything remotely similar to the AMDGPU+Mesa experience on linux. Until then, hacking optimus and unstable kernel updates are the norm.

> I understand that pretty much the only difference between $3000+ "professional" GPUs that get certified for AutoCAD and other workstation systems are identical to $1000 "gaming" GPUs, just with a different on-card BIOS/firmware and e-fuses. I remember ages ago that people were able to simply re-flash their ATI Radeons into FirePro cards and that unlocked higher performance in some applications.

Yep. I have an old laptop with an Nvidia Quadro FX770M GPU that (hackintoshed) macOS sees as a Geforce 9600M GT and runs the included Geforce drivers quite happily on because the FX 770M so close kin with its consumer counterpart.

In fact, in a twist of irony for years the supposedly-more-stable Quadro Windows drivers had a power state bug with this card that would cause the laptop to bluescreen if the GPU tried to ramp down to an idle state. The only solution was to prevent the GPU from idling or to run a different OS with more generic drivers.

I’m not sure if I got tricked or not, it sounds ridiculous but apparently you have to pay extra to get heated backseats in Teslas. It’s a software thing, so it’s not like you need to install new seats. It works similar to miners - consume electricity and generate heat

Who knows, maybe the seat "is" actually using mining to heat the seats. That would be quite the April fools joke.

Every resistive heating device, be it hot water system; space heater; kettle; oven; stove; heat lamp; etc, could just be a proof-of-work IoT device that generates (very little) income.

If everything did that, the profit from doing it would be almost nothing, as difficulty increased. It would also be wildly more expensive than just a simple piece of nichrome wire, and you also now have to have a cellular module and a significant amount of continuous data traffic over it.

Yes, but people would make money in the gradient in between when nothing does it, and when everything does it.

Flashing Radeon 6950 cards with 6970 firmware was commonplace in the 2012 era

That was before the era of VBIOS-signing. You straight-up can't boot any VBIOS image that's not signed anymore.

It was also before the era when PROM fuses were used to permanently disable hardware at the factory. So like, nowadays when you do something like flash Vega 56 with Vega 64 firmware, or flash 5700 with 5700XT firmware, it doesn't actually turn on any cores anymore, they're permanently disabled now. What changes is things like memory speed (Vega) or power limits (5700) that are used to artificially gimp the cards and increase the performance difference beyond the actual core count, as just disabling cores often doesn't give enough difference anymore to be marketable (Vega especially was very bottlenecked by other parts of the card other than the cores).

2012 :-) how about 1999 3Dfx Velocity 1000 TMU unlock to Voodoo3 2000? https://www.anandtech.com/show/401/3

then there was 2003 Radeon 128MB 9500 to 9700 mod https://forum.beyond3d.com/threads/turn-your-radeon-9500-int...

That's not really true anymore.

One of the big differences is RAM.. the A6000 has 48GB, a 3090 only 24 (And the 3080 12). Also, the RAM in the A6000 is ECC, and of course in the consumer cards it isn't.

The A6000 also has ~10% more CUDA and Tensor cores.

So, no, it isn't just a different firmware.

10% cores sounds like softlocked or lasered off cores, rather than a physically different card (unless it’s a single gpu with 10% more cores than a dual GPU).

The differences you're listing off are all matters of how the same underlying silicon is configured during the final stages of manufacturing. The workstation cards that support ECC RAM are using the same RAM chips as the consumer cards, and the memory controller is just configured to use a portion of the memory for ECC instead of as additional capacity. Having one, two or four DRAM dies per memory channel doesn't have any direct connection to whether it's a workstation/server GPU or a consumer GPU, and the consumer GPUs could just as easily be equipped with the same 48GB instead of 12GB or 24GB, if there were sufficient demand.

When you look at the workstation cards that aren't top of the line, it's clear that they are nearly identical hardware to more mainstream consumer gaming GPUs, and the most significant hardware differences are usually smaller coolers on the workstation cards leading to lower clock speeds.

Note: it's less about demand and more about price discrimination. Prosumers realize a higher benefit from their graphics cards and therefore have a higher willingness to pay, but they have higher RAM requirements than do others. This allows nvidia to segment their customer base by restricting the RAM on their gaming cards. If they didn't restrict the RAM on their gaming cards then they couldn't segment them. That doesn't mean there's not consumer demand for cards with higher RAM and marginally higher prices.

Starting with 3000 series the consumer cards use ECC too right?

Newer generations of DRAM (both GDDRx and the DDRx used in desktops and servers) are dense enough that they need on-die ECC. But that only offers some protection for data at rest, while the traditional kind of ECC involves also transmitting extra bits in order to also protect against transmission errors, providing end-to-end data protection. I don't think Nvidia has started allowing the latter kind of ECC to be enabled on consumer products.

Sounds like it isn't exactly the same, but they are able to detect some transmission errors:

https://www.techpowerup.com/review/nvidia-geforce-rtx-3080-f...

That surprises me because memory corruption in something like texture-memory for game world rendering is far-less important than memory corruption in shader program memory. GPU memory needs to be fast, and ECC can slow it down - and add cost. So if they were going to use ECC in gaming GPUs I’d expect it to be limited to areas reserved for shader programs and “program state” rather than output texture memory. If all of VRAM is ECC backed then I’m wowed and kudos to NVIDIA for doing that.

ECC slows down RAM however RAM speeds are starting to become bounded by error rates. To push into the faster timings ECC actually becomes essential to keep the system from becoming unstable.

This is the reason why DDR5 is so much faster despite using ECC effectively by default.

For GPUs, it's not just the RAM timings and clock speeds that matter. The way they implement ECC also directly subtracts from the effective memory bus width and therefore the total bandwidth.

This is true for off-die ECC (i.e the kind that computes ECC between chips using a separate chip) however on-die ECC (the ECC logic and parity memory being included in the memory chip itself) like what DDR5 will have doesn't have this bandwidth issue. It causes the dies to cost slightly more but otherwise largely results in the same performance as non-ECC memory.

In the GPU space, this is why HBM2 memory doesn't really have a performance difference between ECC and non-ECC memory.

Or Infrared cameras, they have the same image sensor, but cost much much more if you by one with a higher resolution (enabled)

https://www.eevblog.com/forum/thermal-imaging/flir-e4-therma...

Oh, like RTX Voice - which 'requires' a 20-series GPU as a minimum requirement. Of course after editing the installer it works just fine on pretty much any GPU with cuda cores.

I’m 100% sure RTX Voice could run happily on a cpu. Might be a fun target for one of the groups working on CPU back ends. It’s a really good NR though; I’ll give it that.

https://krisp.ai/ is basically that, but I've found RTX voice running on the GPU (1660ti) doesn't slow down my laptop while Krisp does a bit.

Marketing needs features to sell the higher end products on. Its easier to bullshit it and artificially limit than to come up with things that physically can only work on the high end sku.

Did anyone ever really try to pitch that RTX Voice actually needed 20 series hardware to work as opposed to being something Nvidia were gating to try to sell more cards? I thought it was pretty transparent myself but maybe I just assumed.

I'm pretty sure this is exactly the reason why Nvidia is entirely unwilling to cooperate on FOSS drivers, and why neither Nvidia nor AMD are willing to support FOSS device firmware.

Which is a shame, because a GPU with FOSS firmware and comparable specs to the AMD/Nvidia hotness could use that software-driven flexibility to its advantage.

Reminds me of how every modern silicon has a "cutting edge, future technology" neural network built into.

Then you notice the asterisk and realize it's just an accelerator (usually just a few instructions with some dedicated logic) and all the NN is still in software.

And you notice the NN software doesn't even use the accelerator half of the time

Raytracing (RTX) is largely a software thing, although real hardware got added to improve ray-triangle and ray-AABB intersection performance.

> real hardware got added to improve ray-triangle and ray-AABB intersection performance.

Isn't that the whole point? I was doing raytracing in the CPU in 2002 for a university project. Doesn't mean I could use it in a game.

Everything is a software thing until it's accelerated by fixed function hardware like, you know, GPUs.

It’s hardware if it’s embedded with C! Never forget that -

With love,

USPTO

You meant NVidia Quadro/Titan workstation performance?

Reply on news.ycombinator.com