Hacker News Reader

Comment by karlkloss | Hacker News Reader

And when the AI bubble bursts, "refurbished" HDDs and GPUs will flood the market. Save your money now and be prepared.

GPUs yes, but there'll be no HDDs making it alive, they'll get destroyed to protect whatever rubbish they had on.

Depends how it goes down, if a company goes into insolvency all security policies are off the table and random hardware can get shifted into lot bidding.

HDD can be written multiple times with random data if data centers really have to protect what their former customers wrote on them. I never looked at those details in standard contracts.

There is also encryption at rest.

> HDD can be written multiple times with random data

Which costs more in compute than simply throwing the drive in a shredder

What's the disposal method for shredded drives?

Mount Doom

Not really, if we give the HDD some resale value. There's a market for used but functional hard drives.

They have enough investor money they don't need to recoup it selling used drives. Straight to the shredder.

All the large datacenter/cloud companies do not let hard drives leave the building.

All you really need to do is write one pass of zeros on them. That will prevent anyone but a very dedicated adversary with expensive equipment from recovering any data, especially on TB scale drives.

Can still take hours per drive though, which is why a lot of people skip it.

I make a random 1MB chunk, then write that all over the drive, at overlapping offsets. I've been told that really clears it. On IDE-spinning-rust disks I trusted it, not sure if I should trust these modern SSD

Why rewrite the same 1MB chunk, instead of making new random chunks?

Redundant data at least opens the possibility that the drive could deduplicate.

Cause making new random was taking too much time.

Pretty sure most modern computers can generate random data faster than they can write to a disc.

Or encrypt it and just trash the encryption header.

When I used to do computer refurbishment, 'Boot and Nuke' was great for this. Load it up at boot, and write over the with random junk a few times.

I wrote hddrand to write random data and optionally read it back to verify integrity. https://github.com/mqudsi/hddrand

I believe many enterprise drives have instant-erase functionality (presumably deleting an encryption key).

If they were encrypted to begin with, yes. Many are not.

I was under the impression that these drives may be transparently encrypted by default. (Rollable encryption key in hardware, invisible to end-user.)

A drive that supports Secure Instant Erase should be encrypting all data. When the SEI function is invoked (“nvme format -s 2”, “hdparm —-security-erase”) they key is thrown away and replaced with a new one. Similar implementations exist for NVMe, SATA, and SAS drives — regardless of whether they are HDD or SSD.

This puts a fair amount of trust and in the drive’s ability to really delete the old key.

Nah, the liquidators aren't going to care about that. Those hard drives are going to be shipped out with all your wildest porn chat bot fantasies.

"Shredded onsite" means by the next user when they format the drive and write contents to it /s

new HDD have a SHA key built in that can be erased without zeroing out the drive. it just creates a new key and boom, free disk

My understanding is that refurbished GPUs from e.g. crypto mining are something people are warned away from, because they've often been run into the ground. Are AI usage practices different?

> My understanding is that refurbished GPUs from e.g. crypto mining are something people are warned away from, because they've often been run into the ground.

No, this wasn't the case. While there were never comprehensive studies various tech media purchased these cards to run testing and found that, other than scammers, they all performed to expectation.

You can buy used car tires with 1% of thread left and they'll perform amazingly during your one time test too

Car tire is a moving part that wears in use. Only moving parts a GPU has are its fans, which can be replaced. The rest may last really long, or fail brand new.

Thermal cycling will eventually cause problems with continuity but I'm not sure that's really predictable. We have some really old GPUs that eventually had problems under load, we attributed it to thermal degradation as we had ruled out pretty much everything else.

Believe it or not but the "really long time" you're mentioning is much less "long" when you run the card 24/7 at 100% in enclosed spaces

Electricity is a moving part btw, electronic components do wear and tear

It was considered a fear but I don't know if there is much truth to it. The fans and capacitors would give out long before the silicon.

Even if it say, halved the life span of the chips, that is still far longer than what most people would ever use them for.

> And when the AI bubble bursts, "refurbished" HDDs and GPUs will flood the market.

GPUs? No way. The datacenter cards don't even have video output ports, and I think the chips destined for AI / ML training also have everything video/render related removed from the silicon, makes for more yield.

And the other way around, using (cheap) consumer GPUs in servers, I think at least NVDA tries to prevent that with driver-based DRM, so there won't be any flooding coming from there either.

If this bubble pops you might need that money for food when bananas go from $1.50 to $150.00

Reply on news.ycombinator.com