Comment by userbinator
4 days ago
One key point about retention which is not often mentioned, and indeed neither does this article, is that retention is inversely proportional to program/erase cycles and decreases exponentially with increasing temperature. Hence why retention specs are usually X amount of time after Y cycles at Z temperature. Even a QLC SSD that has only been written to once, and kept in a freezer at -40, may hold data for several decades.
Manufacturers have been playing this game with DWPD/TBW numbers too --- by reducing the retention spec, they can advertise a drive as having a higher endurance with the exact same flash. But if you compare the numbers over the years, it's clear that NAND flash has gotten significantly worse; the only thing that has gone up, multiplicatively, is capacity, while endurance and rentention have both gone down by a few orders of magnitude.
For a long time, 10 years after 100K cycles was the gold standard of SLC flash.
Now we are down to several months after less than 1K cycles for QLC.
I'm sad that drives don't have a 'shutdown' command which writes a few extra bytes of ECC data per page into otherwise empty flash cells.
It turns out that a few extra bytes can turn a 1 year endurance into a 100 year endurance.
There are programs with which you can add any desired amount of redundancy to your backup archives, so that they would survive corruption that does not affect a greater amount of data than the added redundancy.
For instance, on Linux there is par2cmdline. For all my backups, I create pax archives, which are then compressed, then encrypted, then expanded with par2create, then aggregated again in a single pax file (the legacy tar file formats are not good for faithfully storing all metadata of modern file systems and each kind of tar program may have proprietary non-portable extensions to handle this, therefore I use only the pax file format).
Besides that, important data should be replicated and stored on 2 or even 3 SSDs/HDDs/tapes, which should preferably be stored themselves in different locations.
Unfortunately some SSD controllers plainly refuse to read data they consider corrupted, even if you have extra parity that could potentially restore corrupted data, your entire drive might refuse to read.
22 replies →
Thank you for this.
I had no knowledge of pax, or that par was an open standard, and I care about what they help with. Going to switch over to using both in my backups.
2 replies →
This is fine, but I'd prefer an option to transparently add parity bits to the drive, even if it means losing access to capacity.
Personally, I keep backups of critical data on a platter disk NAS, so I'm not concerned about losing critical data off of an SSD. However, I did recently have to reinstall Windows on a computer because of a randomly corrupted system file. Which is something this feature would have prevented.
Blind question with no attempt to look it up: why don't filesystems do this? It won't work for most boot code but that is relatively easy to fix by plugging it in somewhere else.
Wrong layer.
SSDs know which blocks have been written to a lot, have been giving a lot of read errors before etc., and often even have heterogeneous storages (such as a bit of SLC for burst writing next to a bunch of MLC for density).
They can spend ECC bits much more efficiently with that information than a file system ever could, which usually sees the storage as a flat, linear array of blocks.
7 replies →
You can still do this for boot code if the error isn't significant enough to make all of the boot fail. The "fixing it by plugging it in somewhere else" could then also be simple enough to the point of being fully automated.
ZFS has "copies=2", but iirc there are no filesystems with support for single disk erasure codes, which is a huge shame because these can be several orders of magnitude more robust compared to a simple copy for the same space.
1 reply →
The filesystem doesn't have access to the right existing ECC data to be able to add a few bytes to do the job. It would need to store a whole extra copy.
There are potentially ways a filesystem could use heirarchical ECC to just store a small percentage extra, but it would be far from theoretically optimal and rely on the fact just a few logical blocks of the drive become unreadable, and those logical blocks aren't correlated in write time (which I imagine isn't true for most ssd firmware).
8 replies →
You can, but only if your CPU is directly connected to a flash chip with no controller in the way. Linux calls it the mtd subsystem (memory technology device).
That does sound like a good idea (even if I’m sure some very smart people know why it would be a bad idea)
I guess the only way to do this today is with a raid array?
Because no one is willing to pay for SLC.
Those QLC NAND chips? Pretty much all of them have an "SLC mode", which treats each cell as 1 bit, and increases both write speeds and reliability massively. But who wants to have 4 times less capacity for the same price?
4 times less capacity but 100x or more endurance or retention at the same price looks like a great deal to me. Alternatively: do you want to have 4x more capacity at 1/100th the reliability?
Plenty of people would be willing to pay for SLC mode. There is an unofficial firmware hack that enables it: https://news.ycombinator.com/item?id=40405578
1TB QLC SSDs are <$100 now. If the industry was sane, we would have 1TB SLC SSDs for less than $400, or 256GB ones for <$100, and in fact SLC requires less ECC and can function with simpler (cheaper, less buggy, faster) firmware and controllers.
But why won't the manufacturers let you choose? The real answer is clearly planned obsolescence.
I have an old SLC USB drive which is only 512MB, but it's nearly 20 years old and some of the very first files I wrote to it are still intact (I last checked several months ago, and don't expect it's changed since then.) It has probably had a few hundred full-drive-writes over the years --- well worn-out by modern QLC/TLC standards, but barely-broken-in for SLC.
The real answer is: no one actually cares.
Very few people have the technical understanding required to make such a choice. And of those, fewer people still would actually pick SLC over QLC.
At the same time: a lot of people would, if facing a choice between a $50 1TB SSD and a $40 1TB SSD, pick the latter. So there's a big incentive to optimize on cost, and not a lot of incentive to optimize on anything else.
This "SLC only" mode exists in the firmware for the sake of a few very specific customers with very specific needs - the few B2B customers that are actually willing to pay that fee. And they don't get the $50 1TB SSD with a settings bit flipped - they pay a lot more, and with that, they get better QC, a better grade of NAND flash chips, extended thermal envelopes, performance guarantees, etc.
Most drives out there just use this "SLC" mode for caches, "hot spot" data and internal needs.
2 replies →
> But why won't the manufacturers let you choose? The real answer is clearly planned obsolescence.
No, it's not. The real answer is that customers (Even B2B) are extremely price sensitive.
Look, I know the prevailing view is that lower quality is some evil corporate plan to get you to purchase replacements on a more frequent basis, but the real truth is that consumers are price sensitive, short sighted, and often purchasing without full knowledge. There's a race to the bottom on price, which means quality suffers. You put your typical customer in front of two blenders at the appliance store, one is $20 and the other is $50, most customers will pick the $20 one, even when armed with the knowledge that the $50 version will last longer.
When it comes to QLC vs SLC, buyers don't care. They just want the maximum storage for the smallest price.
1 reply →
> I have an old SLC USB drive which is only 512MB, but it's nearly 20 years old and some of the very first files I wrote to it are still intact (I last checked several months ago
It's not about age of drive. It's how much time it spent without power.
1 reply →
Funny enough I just managed to find this exact post and comment on google 5 minutes ago when I started wondering whatever it's actually possible to use 1/4 of capacity in SLC mode.
Though what make me wonder is that some reviews of modern SSDs certainly mention that that pSCL is somewhat less than 25% of capacity, like 400GB pSLC cache for 2TB SSD:
https://www.tomshardware.com/pc-components/ssds/crucial-p310...
So you get more like 20% of SLC capacity at least on some SSDs
> do you want to have 4x more capacity at 1/100th the reliability?
Yes.
QLC SSDs are reliable enough for my day-to-day use, but even QLC storage is quite expensive and I wouldn't want to pay 4x (or realistically, way more than 4x) to get 2TB SLC M.2 drives instead of 2TB QLC M.2 drives.
NVMe protocol introduced namespaces. Is it not the feature perfect for users to decide themselves, how to create 2 virtual SSDs with TLC and pseudo-SLC-mode, choosing how much space to sacrifice for pSLC?
1 reply →
> Alternatively: do you want to have 4x more capacity at 1/100th the reliability?
If the original drive has sufficient reliability, then yes I do want that.
And the majority of consumers do, too.
Chasing absolute extreme highest powered off durability is not a priority for 99% of people when the drives work properly for typical use cases. I have 5 year old SSDs where the wear data is still in the single digit percentages despite what I consider moderately heavy use.
> I have an old SLC USB drive which is only 512MB, but it's nearly 20 years old and some of the very first files I wrote to it are still intact (I last checked several months ago, and don't expect it's changed since then.) It has probably had a few hundred full-drive-writes over the years --- well worn-out by modern QLC/TLC standards, but barely-broken-in for SLC.
Barely broken in, but also only 512MB, very slow, and virtually useless by modern standards. The only positive is that the files are still intact on that old drive you dusted off.
This is why the market doesn’t care and why manufacturers are shipping TLC and QLC: They aren’t doing a planned obsolescence conspiracy. They know that 20 years from now or even 10 years from now that drive is going to be so outdated that you can get a faster, bigger new one for pocket change.
1 reply →
> If the industry was sane
Industry is sane in both the common and capitalist sense.
The year 2025 and people still buy 256Tb USB thumbdrives for $30, because nobody cares except for the price.
To be honest you can buy 4TB SSD for $200 now, so I guess market would be larger if people were aware of how easy would it be to make such SSDs work in SLC mode exclusively.
Myself wants. I remember when the UBIFS module (or some kernel settings) for the Debian kernel was MLC against SLC. You could store 4X more data now, but at a cost of really bad reability: A SINGLE bad shutdown and your partitions would be corrupted up to the point of not being able to properly boot any more, having to reflash the NAND.
Well then buy an industrial SSD, they're something like 80-240 GB and you get power loss protection capacitors too. Just not the datacenter ones, those melt immediately without rack airflow.
It is not 4 times, it is 16 times. Your 1TB storages would become 64GB.
Wrong? You are comparing number of possible values that can be stored not the amount of bits. Drive sizes are specified in bytes not possible combinations it can store.
Endurance going down is hardly a surprise given that the feature size has gone down too. The same goes for logic and DRAM memory.
I suspect that 2035 years time, hardware from 2010 will work, while that from 2020 will be less reliable.
Completely anecdotal, and mostly unrelated, but my NES from 1990 is still going strong. Two PS3’s that I have owned simply broke.
CRTs from 1994 and 2002 still going strong. LCD tvs from 2012 and 2022 just went kaput for no reason.
Old hardware rocks.
LCD tvs from 2012 and 2022 just went kaput for no reason.
Most likely bad capacitors. The https://en.wikipedia.org/wiki/Capacitor_plague may have passed, but electrolytic capacitors are still the major life-limiting component in electronics.
3 replies →
Specifically old Japanese hardware from the 80s and 90s - this stuff is bulletproof
1 reply →
I had an LCD that worked from around 2005 to 2022. It became very yellow closer to 2022 for some reason. It was Samsung PVA, I think it was model 910T.
2 replies →
For what it's worth my LCD monitor from 2010 is doing well. I think the power supplied died at one point but I already had a laptop supply to replace it with.
I concur; in my experience ALL my 24/7 drives from 2009-2013 still work today and ALL my 2014+ are dead, started dying after 5 years, last one died 9 years later. Around 10 drives in each group. All older drives are below 100GB (SLC) all never are above 200GB (MLC). I reverted back to older drives for all my machines in 2021 after scoring 30x unused X25-E on ebay.
The only MLC I use today are Samsungs best industrial drives and they work sort of... but no promises. And SanDisc SD cards that if you buy the cheapest ones last a surprising amount of time. 32GB lasted 11-12 years for me. Now I mostly install 500GB-1TB ones (recently = only been running for 2-3 years) after installing some 200-400GB ones that work still after 7 years.
> in my experience ALL my 24/7 drives from 2009-2013 still work today and ALL my 2014+ are dead,
As a counter anecdote, I have a lot of SSDs from the late 2010s that are still going strong, but I lost some early SSD drives to mysterious and unexpected failures (not near the wear-out level).
1 reply →
As far as I'm aware flash got a bit of a size boost when it went 3D and hasn't shrunk much since then. If you use the same number of bits per cell, I don't know if I would expect 2010 and 2020 or 2025 flash to vary much in endurance.
For logic and DRAM the biggest factors are how far they're being pushed with voltage and heat, which is a thing that trends back and forth over the years. So I could see that go either way.
Oh, it would be nice if it were just feature size. Over the prior 15 years, the nand industry has doubled its logical density three times over with the trick of encoding more than one bit per physical voltage well, making the error bounds on leaking wells tighter and tighter and amplifying the bit rot impact, in number of ECC corrections consumed, per leaked voltage well.
I recently found a 1GB USB drive from around 2006 I used to use. I plugged in and most of the files were still readable! There were some that were corrupted and unreadable unfortunately.
I also seem to remember reading retention is proportional to temperature at time of write. Ie, best case scenario = write data when drive is hot, and store in freezer. Would be happy if someone can confirm or deny this.
I know we're talking theoretical optimums here, but: don't put your SSDs in the freezer. Water ingress because of condensation will kill your data much quicker than NAND bit rot at room temperature.
I'm interested in why SSDs would struggle with condensation. What aspect of the design is prone to issues? I routinely repair old computer boards, replace leaky capacitors, that sort of thing, and have cleaned boards with IPA and rinsed in tap water without any issues to anything for many years.
1 reply →
Would an airtight container and liberal addition of dessicants help?
4 replies →
What about magnetic tape?
1 reply →
That's probably this: https://www.sciencedirect.com/science/article/abs/pii/S00262...
I definitely remember seeing exactly this.
That's how it has to work. To increase capacity you have to make smaller cells where charge may easier diffuse from one cell to another. Also to make drive faster, stored charge has to be smaller, which also decrease endurance. With SLC and QLC comparison is even worse as QLC is basically clever hack to store 4 times more data in the same number physical cells - it's tradeoff.
Yes, but that tradeoff comes with a hidden cost: complexity!
I much rather have 64GB of SLC at 100K WpB than 4TB of MLC at less than 10K WpB.
The spread functions that move bits around to even the writes or caches will also fail.
The best compromise is of course to use both kinds for different purposes: SLC for small main OS (that will inevitably have logs and other writes) and MLC for slowly changing large data like a user database or files.
The problem is now you cannot choose because the factories/machines that make SLC are all gone.
The problem is now you cannot choose because the factories/machines that make SLC are all gone.
You can still get pure SLC flash in smaller sizes, or use TLC/QLC in SLC mode.
I much rather have 64GB of SLC at 100K WpB than 4TB of MLC at less than 10K WpB.
It's more like 1TB of SLC vs. 3TB of TLC or 4TB of QLC. All three take the same die area, but the SLC will last a few orders of magnitude longer.
1 reply →
My problem is: I have more than 64GB of data
On the other hand when capacity goes up, the cycle-count goes down for the same workload. A 4TB drive after 1K cycles has written the same amount of data as 100GB drive after 40K cycles.
> Even a QLC SSD that has only been written to once, and kept in a freezer at -40, may hold data for several decades.
So literally put your data in cold storage.
That is literally the origin of the term.
So AWS S3 Glacier might actually be cold