Using voltage glitching to unleash the A100 in my mining card

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.

6 months ago I looked at used rtx3090s. I marvelled they cost a little over what I paid over a year ago. I ordered 2 more for $900 if I remember correctly.

3 months ago the cheapest used one was $1300, a month ago you could get it for $1500 now it is $1650.

I thought the prices would drop once they reach the peak. But the peak seems to be far away.

A 5090 is only twice faster in typical inference, but costs $10k now. Insanity.

The highs and lows of my predictive capabilities

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 not cheap, so features/bugs like this can remain around for a while.

In both IC companies I worked for, had around 6 ICs. If you opened any of them up, you would see the exact same die in every one. After a couple years of great yield, you would then find half missing in the lowest tier.

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).

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

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.

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.

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.

All sorts of reasonable interactions are then policed by people with visceral reactions to normal activity.

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.

That said, 80G for something that you picked up for cheap could be an absolute steal these days.

Bad RAM isn't a huge issue as you can flag and ignore the relevant regions and be down a few MB.

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.

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.

I think my story may make your grapes less sour.

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.

As for v100, p100 etc I considered them, but they make zero sense as they don't have int8/int4. Everything gets upscaled in them which slows the inference down.

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 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.

Hacking is a good thing.

If you wanted to patch a die the only real means is through a FIB (Focused Ion Beam) workstation, and even then you need the die structure to allow it (you have to mill your way into the chip and can't have any functional layers in between you and the patch you want to makey), and even when doable it would be an insanely expensive, time-consuming and likely fragile project.