Anecdotically, 64GB works extremely reliably on 8GB cards after unlock.
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".
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.
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.
On the other hand ... perhaps a few bit errors is not a problem for AI workloads ... but they could also crash your system ...
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?
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.
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).
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.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).
I usually miss out on these things, but I'd been monitoring the eBay prices rising for inferior cards and figured this was a good bet.
Annoying thing is finding a good motherboard to put multiple blower cards in. Mining rigs are a steal but PCIe x1 limitation is annoying. Got a bunch of Volta cards too sitting around until I figure out where to put them (and the performance per watt sucks next to the 170).
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.
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.
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 wouldn't really work for AMD processors... They could design and fab a 6-core chiplet, but it would need to be the same size as the 8-core chiplets so it would interface to the same carriers. If the footprint is the same, the fab cost is roughly the same, so spending the design time on that doesn't make sense. If they build separate carriers, that's more design time and inventory management and it may well be more expensive than just building 8-cores and disabling two.
That said, with Zen6 having 12 cores per chiplet, I wouldn't be surprised if they eventually end up designing a smaller chiplet with 6 cores and a dedicated small chiplet carrier. Or push 6 core SKUs to the APU side that doesn't use chiplets, Zen5 APUs already have specialized masks for some of their SKUs.
Its not economies of scale. Why or how would you even etch a worse pattern on the same fabrication node? They don’t even have a mask for the worse version its literally just the same pattern with more defects.
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.
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.
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.
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.
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.
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.
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.
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.
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.
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.
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.
Anyways, that's what I tell myself as I taste my bowl of sour grapes.
I remember looking at research n thinking if they'd be useful for AI with their integer ops
I even looked into blog posts of using them or unlocking them somehow but I never thought to think that they actually had the full HBM memory and all just fused away??? wow
> 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!
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.
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.
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.
> 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 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.
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
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.
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".
Those were binned cards, though.
But what if nvidia used those fuses because an A100 had too many fabrication faults so it could only be a cheaper mining card?
(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.
But nowadays they can make much more money with enterprise targeted products.
Which is why I think these devices are more likely to be faulty.
The author mentions about 90% yield on getting the full 80GB of memory (in a limited 10 card sample).
Nowdays - but this was a mining card, meaning bought for the crypto boom and before AI.
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.
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.
On the other hand ... perhaps a few bit errors is not a problem for AI workloads ... but they could also crash your system ...
They say ECC works (?) so you would know.
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.
HBM is so expensive that I doubt you would find significant gains on average. They are already yieldmaxxing like crazy.
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?
we call that dithering in the mechanical and analog worlds
The sheer recklessness on display here. Everyone knows random bits is how something becomes sentient! You'll kill us all!
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
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).
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.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).
Great usage of glitching!
I just wish they had something like this for older Tesla p100s, p4s...
There was a Pascal era mining card that has the same double your ram trick https://github.com/6Robert123/cmp100-100-unlock
Glad I jumped on these cards, excited to give this a try!
Hard to believe $1k was a deal a few months ago (and $200 before then).
I really wish I did. I told a friend about how I bet an unlock would come and decided to buy them when I heard about an unlock.
The highs and lows of my predictive capabilities
I usually miss out on these things, but I'd been monitoring the eBay prices rising for inferior cards and figured this was a good bet.
Annoying thing is finding a good motherboard to put multiple blower cards in. Mining rigs are a steal but PCIe x1 limitation is annoying. Got a bunch of Volta cards too sitting around until I figure out where to put them (and the performance per watt sucks next to the 170).
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.
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.
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.
When yields get better, manufacturer can just produce the lower tier chip directly.
This wouldn't really work for AMD processors... They could design and fab a 6-core chiplet, but it would need to be the same size as the 8-core chiplets so it would interface to the same carriers. If the footprint is the same, the fab cost is roughly the same, so spending the design time on that doesn't make sense. If they build separate carriers, that's more design time and inventory management and it may well be more expensive than just building 8-cores and disabling two.
That said, with Zen6 having 12 cores per chiplet, I wouldn't be surprised if they eventually end up designing a smaller chiplet with 6 cores and a dedicated small chiplet carrier. Or push 6 core SKUs to the APU side that doesn't use chiplets, Zen5 APUs already have specialized masks for some of their SKUs.
They still wouldn't because economies of scale make it cheapest to only build one model.
Its not economies of scale. Why or how would you even etch a worse pattern on the same fabrication node? They don’t even have a mask for the worse version its literally just the same pattern with more defects.
They only make one design because of economies of scale. I think there might have been some confusion here
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.
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.
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.
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.
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.
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.
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.
So I guess the next step is to stop selling the lower-binned card? https://www.tomshardware.com/pc-components/gpus/nvidia-repor...
0: https://news.ycombinator.com/item?id=50452839
Prepare to be repugnantated regularly.
With a bit of nuance of understanding it's not that wild or egregious, and no one is really getting cheated.
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.
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.
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.
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.
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.
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.
That said, 80G for something that you picked up for cheap could be an absolute steal these days.
No need for underclocking, though power limiting doesn't sacrifice much performance.
Currently listening to the soothing drone of my $200 Octominer cooling my cards while I'm fine tuning a model (almost makes a good space heater).
I run mine in octominer too, with fan control too. No noise, did fine during the winter with heat and it's very comforting as we enter winter.
https://www.reddit.com/r/LocalLLaMA/comments/1wjr59t/768gb_v...
Yeah the fan noise is only noticeable when I push the cards to ~180W, otherwise they idle pretty quiet.
How are you dealing with the RAM limit (32GB)? I had upgraded mine to an i7 but it was unstable so I’m back to the stock Celeron class CPU.
I sensed the sour grapes :D
Bad RAM isn't a huge issue as you can flag and ignore the relevant regions and be down a few MB.
I remember looking at research n thinking if they'd be useful for AI with their integer ops
I even looked into blog posts of using them or unlocking them somehow but I never thought to think that they actually had the full HBM memory and all just fused away??? wow
Finally, I can download more VRAM!
Just don’t download a car.
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!
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.
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.
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
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
> 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 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?
Joke aside, good investigation and blog debut, kate!