From looking at numbers from the macOS SSD thrashing issue, we know lifetime usage as reported by SMART is not a strict function of terabytes written (at least on those SSDs). This makes sense, because small writes have a much higher endurance cost than large writes on SSDs. SSD lifetime isn't based on total bytes written, it's based on total blocks erased in the underlying Flash, and blocks are much larger than sectors so small writes necessarily cause write amplification to some extent (depending on the wear leveling).

If manufacturers are guaranteeing a certain number of TBW on SSDs but can't actually sustain that at pathological write sizes, they need to start qualifying those guarantees with e.g. an average write size statistic. Then they can claim Chia breaks that assumption and reject warranty claims in those cases.

This isn't unique to Chia. For example, a database workload will cause higher stress per byte written on an SSD than a raw video recording workload.

It's also possible they're talking about time-based warranties with daily usage numbers. Many SSDs will be "guaranteed for 5 years" at some DWPD (drive writes per day). In that case they're probably just saying Chia plotting greatly exceeds the DWPD stat and thus your number of years warranty no longer applies. That ignores the above problem and is just about how the drive endurance is specified.

(Also, just to clarify: apparently SSDs are used in Chia as a staging ground to build data to put on HDDs, which is then read only; I don't think there is any endurance cost to disks for actually "farming" Chia, only for "plotting" the contents of those disks initially; I think Chia is silly but just want to get the tech right)

Is there any way to extend the lifetime of a SSD by allowing its capacity to decrease by replenishing its reserve of replacement blocks so that the SSD could continue to operate until it runs out of writable blocks?

A SSD that presents as a traditional hard-drive style block storage device can't shrink itself on the fly without causing catastrophic software problems. The least bad way to implement such behavior would be for the highest-numbered LBAs to be removed first, which means your backup partition table is first to disappear, then you start encountering filesystem corruption.

A zoned SSD has an explicit mechanism for the drive retiring zones from service and making them read-only or completely inaccessible. But it's not currently practical to use a zoned storage device as the primary or sole storage in a system; at a minimum you need a regular block device or namespace to boot an OS from.

Theoretically this would be possible but probably wouldn't help much in practice. Writes on an SSD aren't mapped 1:1 from "logical" address to physical address - they constantly balance and move writes around the entire device (the "Flash Translation Layer" it's sometimes called). As a result once you have exhausted the normal reserve capacity (perhaps allowing for some early life failures) I suspect it's likely the rest of the device will soon follow and you wouldn't extend it's life by that useful an amount.

I don't have hard data however on like the "curve" of flash cell failures as an SSD ages and therefor exactly how much longer you might get based on the minimum to median cell failure.

Generally, you would look to a datacenter line of SSD which is rated for much higher endurance. These are typically overprovisioned much more than consumer ssds in order to meet the endurance requirements.

I think most OEMs opt not to let you choose this sort of thing and rather bury it into the drive firmware.

Does Chia do lots of small writes? A lot of these plot generation steps sound like the writes are pretty linear. Plus sorting, which can easily be designed to use big writes too.

That sounds like what they're saying. I imagine when they determine warranties they're looking at what they estimate to be "normal" use? Reading/writing to capacity 24/7 certainly creates a lot of heat for a long time.

Its not the heat. SSDs wear out after repeated writing. Some more quickly than others.

That was already known, it's an included spec when you're buying an SSD and already written into the warranties... this is an additional warning for this use case:

>“If users use our SSDs for mining/farming and other abnormal operations, the data writing volume is much higher than the standard for daily use, and the SSD will slow down or get damaged due to excessive data writing volume. Due to the tests carried out, the damages are qualitative according to the test results, and that is why according to the quality assurance standards of our SSDs, we have the right to refuse to provide warranty services. The right of final interpretation belongs to the company."

They're specifically calling out "standard for daily use"...

One alternative explanation is that maybe farmers were just wearing them out quickly and playing naive to claim the warranty?

Most SSDs aren't designed for the repeated writing Chia requires. "Data center" or "high-endurance" SSDs have that capacity but are consequently more expensive.

>One alternative explanation is that maybe farmers were just wearing them out quickly and playing naive to claim the warranty?

That's exactly what they were doing. If you use a low cost "consumer" SSD you can wear it out very quickly then claim it was "defective."

> Most SSDs aren't designed for the repeated writing Chia requires.

It doesn't require all that many writes. Problems come from a conscious decision to use a small number of SSDs as scratch space for a very large array.

Is that a false claim?

The expectation that an SSD or even SD-card is infinitely rewriteable (practically speaking) is implicit in the marketing.

The idea that a drive can't be rewritten too many times should be explicit (like SMR HDDs)

Every SSD spec sheet has a Total Bytes Written (TBW) spec and Drive Writes Per Day (DWPD) and many warranties say that they don't cover wearout beyond those specs. It's akin to a tire rated for 50,000 miles; if it fails beyond that it's not defective.