> 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.
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.
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.
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.
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.
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 ...
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.
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 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).
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...
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?
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
> 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.
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)
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.
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.
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.
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.
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 ...
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.
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
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.
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.
I just wish they had something like this for older Tesla p100s, p4s...
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).
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 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
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...
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?
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?