amazing to think there was a time (literally not too long ago) where you would throw 80 GB of HBM onto a chip and turn 90% of it off
Those were binned cards, though.
Binned for compute failures, not memory failures.<p>That's why this works. The yield on those HBM stacks is excellent. The binned cards had failures in the logic die, and hyperscalers want uniformity in their clusters. That's why the mining market took the downbinned cards.
I'm wondering if this is actually the case or whether they were binned for defects in the HBM
Anecdotically, 64GB works extremely reliably on 8GB cards after unlock.<p>It is much more of "let's sell cards to crypto people, but nerf them, to not harm our enterprise margins", not "we've got a ton of subpar hardware, let's to something with it".
Sam Altman is probably buying them all up now
Instead of putting the host to ACPI S3 state, could an external PCIe riser be used with a separate power supply?<p>Seems like that would considerably speed up the process, by letting the host continue to run while only the target cycles.
This is great. I’m actually trying to unlock the 80G in a CMP 170HX the software way for the last 2 weeks (still no luck).<p>I just don’t know if I’m brave enough (and knowledgeable enough) to go this blown fuses way. Seems a lot more risky.<p>P.S. I now see this is for the 8GB model, which is insane. The 10G one is the one that until now was proven to have 80G, but never fully accessible through CUDA (it was accessible using CPU memory copy though - much slower).
But what if nvidia used those fuses because an A100 had too many fabrication faults so it could only be a cheaper mining card?<p>(which is the more likely scenario because they can make much more money on enterprise/datacenter products, see DRAM)
The article mentions that some die areas are factory-marked as defective and have to be "re-disabled" after the unlock. Ctrl-F _DEFECTIVE
Isn’t this how these things are always done? Same with number of cores in a CPU, the higher core and lower core come from the same fabrication process, it’s just that the lower core CPU had more faulty cores that are disabled.
Depends a lot on when its from in the product cycle, the yield at the end of production runs are usually so good that it turns into artificial product segmentation
There's an aspect of that, but the primary reason for differences between various CPUs and GPUs which have the same underlying chip is product segmentation.
Historically, there are products with high enough yield rates that it makes it irrelevant. See ATI Radeon 9500 Pro. A two byte patch in the firmware gives you an ATI Radeon 9800 Pro.
Yes, that's why I question the approach here.<p>Testing hardware is a complicated matter (you can do a PhD on the topic), and nvidia probably does a better job at it than a hacker in their basement who thinks they can upgrade their GPU for free.<p>On the other hand ... perhaps a few bit errors is not a problem for AI workloads ... but they could also crash your system ...
The good news is that Nvidia <i>did</i> test the hardware, and left the "defective" masks right there in the fuses for us to read. If you have a card that has been fuse-segmented well below its defective part count, you have a product segmentation product on your hands.
They say ECC works (?) so you would know.<p>I recently had a faulty A100, for some reason the provider turned off ECC on that card, so it silently failed and gave junk results... Took me a minute to figure out.
Some applications may be more fault-tolerant than others though. If an LLM glitches you can reinitialize it, maybe reshuffle the allocations a bit and it'll be fine?
HBM is so expensive that I doubt you would find significant gains on average. They are already yieldmaxxing like crazy.
They definitely did, but maybe you can get some good results running your own tests and determining if they really ARE bad or not, or they just overly tagged them out.
it appears the opposite is true and that's why it's faster than an A100, which is optimized for yield and allows more defects
Glad I jumped on these cards, excited to give this a try!<p>Hard to believe $1k was a deal a few months ago (and $200 before then).
Considering how unobtainium GPUs have become nvidia could get a PR win by publishing unlocks for all their outdated hardware. ... people are figuring it out anyways, might as well embrace it.
Unlocks and VRam upgrades, like VRam sim slots, and a VRom slot that ran Linux, had its own GUI for reconfiguration. Like old 68k macs that have a selectable os in the boot roms.
I just wish they had something like this for older Tesla p100s, p4s...
Great usage of glitching!
<i>there are several functions where blowing a fuse on an NVIDIA die can grant you more access, not less. In a properly secured design, a blown fuse should always induce a more restrictive state, because a blown fuse cannot be unblown in an economically feasible way.</i><p>Not really.<p>If the fuses are for security, sure. But these aren't.<p>These fuses are for economic extraction. Doing it the way NVIDIA did it is a safety feature: it ensures that blowing a fuse will never destroy economic value -- it will never turn a card that was previously able to do something correctly into a card that can no longer do that thing correctly.
No, that's absolutely not the perspective of the bean counters at hardware companies. The fuses are also for economic security. If someone wants more features, you make sure that they paid for them. You don't give away higher tier SKU <i>functionality</i> for free. There was probably some cost/engineering justification for it, possibly like you say! Maybe too high of yield. This will almost certainly be removed in the next revision, as have all the many other similar cases before it.<p>For those unaware, product tiers come naturally with IC's, by disabling physical sections of the chip that can't work due to inevitable manufacturing defects. You don't throw defective chips away, you layout in a way where you can physically cut off those defective sections of the chip, with a saw and/or with a fuse, in a way that minimizes lost sales. This works great in the beginning, when your yields are low and you're still dialing in the manufacturing process. But, as yields increase, you might not have enough of the lower tier chips to meet demand. So you, just fuse otherwise good chips. Chips are cheap, after all! New revisions are <i>not</i> cheap, so features/bugs like this can remain around for a while.<p>In both IC companies I worked for, had around 6 ICs. If you opened any of them up, you would see the <i>exact</i> same die in every one. After a couple years of great yield, you would then find half missing in the lowest tier.
Though if it enables hardware that wasn't enabled because of binning, making the card unstable, can you go back you go back to having a working card?
Selling anything that has been purposely hamstrung to limit available capabilities for market reasons seems morally repugnant beyond words. I understand walling off die defects but if it is functional that is disgusting.
I understand your position, but as someone who prefers the lower end products (and the lower end prices!), this policy would be detrimental for me.<p>At the beginning of a product, when yields are bad, I could get a low tier chip at a good price, but then as yields get better, low tier chips wouldn't be available and I no longer have a low cost option. Presumably higher yields would lead to reduced prices on the higher tier chips, but I don't think they'd get down to where the low tier chips were.<p>Another alternative is that the maker only ever sells chips that test successfully as high tier, but I think it's repugnant to not make use of parts with defects if the defects can be gated off.
This is one of the reasons a market is superior to centralized taste-setters. Despite the fact that both a seller and buyer can agree on a price for a less capable product, taste-makers often have these strange neuroses about what they should be paying for.<p>All sorts of reasonable interactions are then policed by people with visceral reactions to normal activity.
You’re going to be very disappointed to discover that this is common practice for many technology products. It’s cheaper and less risky to design a small number of things and then segment their capabilities than it is to make custom things for every different market segment.<p>It’s also useful for binning: These lower spec locked products can have defects in the unused parts, like flakey memory regions or channels. Not every CMP170 can be fully usable after unlocks.<p>The alternative to this market segmentation isn’t “Everyone gets the whole thing for the lowest price”. It’s that the lower price version doesn’t exist at all, or if it does it’s a lot more expensive.
It's really no different. They are selling you a product with X advertised specs, and indeed the specs you can access as the customer (at least out of the box) are those exact advertised specs.
As this comment claims[0], it does seem like eventually it gets to where what you're saying is true, in that it's basically artificial segmentation.<p>So I guess the next step is to stop selling the lower-binned card? <a href="https://www.tomshardware.com/pc-components/gpus/nvidia-reportedly-halts-geforce-rtx-5090-production-in-favor-of-ai-data-center-and-professional-gpus-impending-supply-drought-expected-to-drive-up-prices-rtx-5080-24gb-rumored-as-new-gaming-flagship" rel="nofollow">https://www.tomshardware.com/pc-components/gpus/nvidia-repor...</a><p>0: <a href="https://news.ycombinator.com/item?id=50452839">https://news.ycombinator.com/item?id=50452839</a>
Prepare to be repugnantated regularly.
Imagine a world without this, where nothing was hamstrung and there was only one price point instead of a hamstrung low price and a nonhamstrung high price.<p>That one price would be somewhere between the low price and the high price. So, the world you want to live in would be one where "the poor" who currently pay the low price would need to pay more. And "the rich" would get a discount from their high price they pay currently. That seems morally repugnant beyond words.
With a bit of nuance of understanding it's not that wild or egregious, and no one is really getting cheated.<p>Basically the high price of the fully unlocked models is what makes it possible to sell artificially limited models at the lower price point (while also offering the ability to possibly sell lower binned units). Basically you can gas the profit on the high-end to drive down the cost on the low end, but obviously this only works if you have a distinct High and Low end.<p>People think that if we banned this practice it would lead to the top shelf models being sold at the bottom shelf price. But that's just naive market confusion. A ban would just mean that the top shelf unit at a higher price and the lower shelf unit at a higher price would be available.<p>There is no solve to the "problem" that results in "High shelf unit being sold at lower cost", in fact the thing you are complaining about is already the closest to that ideal, as the manufacturer can leverage economies of scale to drive down costs across the product line.
Before you get too excited about the CMP 170HX, the prices have already skyrocketed and that card doesn't have a cooler on it. It is designed for a server cooler that sounds like a jet engine. Also most of the 80GB vram unlocks have bad RAM and that is the reason they were locked in the first place, otherwise they would have became A100s.<p>Anyways, that's what I tell myself as I taste my bowl of sour grapes.
Yeah, cooling AND underclocking will probably be required to somewhat safely operate these.<p>That said, 80G for something that you picked up for cheap could be an absolute steal these days.
Here is where I get to smugpost about how I bought a ton for $800 each, against the disgresion of my friends and claude who did not believe in the exploit...<p>After which I sold most of them for $2000-$2600 (I believed the panic about the new mac studio launch) instead of holding them to sell for like $5k and making my annual income in a week. And also I failed to do the capacitor mod after buying a couple hundred in soldering equipment. But at least I sold most of them for monero, and that's gone up since and allowed me to avoid massive ebay fees.<p>The moral of the story here is that you can fantisize about getting in at the bottom, but even if you did you may not have rode it to the top. Plenty of people bought bitcoin when it was half a cent and sold at the local maxima to net $30k and spent it on a used corrolla.<p>I think my story may make your grapes less sour.<p>Edit: I knew at the time of purchase that the 10GB->80GB unlocks were problematic but the 8GB->64GB cards were solid. The pricing reflects this, so thats just skill issue.
I sensed the sour grapes :D<p>Bad RAM isn't a huge issue as you can flag and ignore the relevant regions and be down a few MB.
Finally, I can download more VRAM!
The ebay auction headlines have already rewritten themselves.
> In one of the most extreme manifestations of product segmentation, the CMP 170HX, an Ethereum mining card with nominally 8 or 10 GB of application-exposed framebuffer capacity, actually comes with, respectively, ~80 and ~40 GB of HBM onboard!<p>wtf, it seems RAM truly was too cheap back then...
The marginal cost to fabricate the GPU die with the full ram allotment, and then disable a portion for 1) salvaging otherwise useless die's with partly bad ram for lower ram use cards; or 2) market segmentation; is tiny compared to the marginal cost of creating X different masks and fabricating X different die's, one specific to each market segment/ram size amount.<p>In addition, fabricating the same die for every SKU point in the product line means the maker does not have to guess how many version X vs. version Y vs. version Z dies to fabricate. The just have one fabricated. Then the dies become SKU version X, Y or Z as demand requires during the card assembly process.
The card was initially priced at over 4 grand and it had its capabilities severely reduced to make it not compete with real GPUs.<p>Plus: most of those cards probably have broken RAM chips anyway. Throwing them out would be a waste, limiting trash chips to a small amount of RAM that does work turns recycling costs into a product that can be sold for money.
Scaling according to demand is way too hard when it comes to silicon.
Nvidia might've been planning to eventually buy them back from the miners, or sell software RAM unlocks.
>an Ethereum mining card<p>Wait, so does this unlock it to actually do something useful, or just "mine" marginally more shas?
Eth isn’t mined anymore so that wouldn’t be particularly useful
You can use it for machine learning, or anything a normal A100 is used for.
Imagine if cars were like this: "We will sell you this Ferrari for $2K and it will perform like a Chevy Vega until you give us another $200K, which will unlock the hardware."<p>Oh wait, cars <i>are</i> like this now.
Swear to God, this is absolutely true. Had a friend who had a RAV4. No tow package for low end torq. So... waited a few months, spotted a programmer on Craigslist, said it had tow mode. It had 4 more models too. Super tow, super economy and stock. Buyer was beside themselves, learned how well they worked, and would switch to super eco on one ramps. ( really a bad idea, so they stopped...)<p>Hacking is a good thing.
It's utterly crazy that anyone outside Nvidia knows any of this, much less all of it. Has there been a years long public reverse engineering effort that I didn't hear about until now?
Yes, there actually was a years long public reverse engineering effort; since 2023 there were some academic papers, solving parts of it. It has sped up a lot recently, probably due to accumulated knowledge, and due to AI use.
If you think about it, the effort is economically incentivized.<p>Buy a shitton of these things, spend a year working on the unlock. If you find one, sell the shitton at 5x higher price (which is what happened).
[flagged]
> In a properly secured design, a blown fuse should always induce a more restrictive state, because a blown fuse cannot be unblown in an economically feasible way.<p>I disagree: If there's any risk that the fuse will blow with normal usage, then a blown fuse means a defective product.
These are not safety fuses. They will never, in normal use, blow. These are fuses used only in a logic circuit. There is a way that the factory can make the fuse blow and then the circuit code or whatever can check is that fuse blowing and if so it will not do something.
What if the author turn out to be Alice Huang?<p>Joke aside, good investigation and blog debut, kate!