Does anyone make good SSDs any more?

[PlainName]

Doesn't wear leveling achieve that? Use half your disk and you've doubled the cells available for endurance.

[mariush]

It depends on the controller

Some can switch the whole free space to pseudo-SLC and only convert to TLC or QLC when nearly all buffer is full. 

Some controllers have a dynamic or static + dynamic portion, up to a percentage of drive's space ... for example some older WD drives like SN570 had a maximum of 12 GB pseudo-slc write cache. Some samsung drives have a guaranteed 6 GB fixed (always there) + up to something like 150-250 GB for every 1 TB or 2 TB of TLC memory.

[David Hess]

Doesn't wear leveling achieve that? Use half your disk and you've doubled the cells available for endurance.

It is not about using less of the drive; all of the cells are available for endurance.  Larger drives have greater total endurance because they have more cells available for wear leveling, and that is one way to get longer operating life, but of course it comes at the cost of buying a drive with more storage.

Drive endurance is proportional to size, but how many bits are stored per cell has an even greater effect, so a drive with the same number of cells but fewer bits per cell should have higher endurance despite lower capacity.

[PlainName]

but of course it comes at the cost of buying a drive with more storage.

Yes, exactly. The question was why don't they sell smaller drives with longer life (due to SLC vs QLC), and my contention was that you might achieve the same thing from just buying a bigger drive and not using some of it, even if it has to be a huge drive in comparison - same usable size, same lifetime, just a different way to get there.

Edit: Of course, that's predicated on unused space being usable for wear leveling (which apparently may not be the case) and wear leveling actually not losing data. But it's a mature technology and cheap as chips, now

[David Hess]

but of course it comes at the cost of buying a drive with more storage.

Yes, exactly. The question was why don't they sell smaller drives with longer life (due to SLC vs QLC), and my contention was that you might achieve the same thing from just buying a bigger drive and not using some of it, even if it has to be a huge drive in comparison - same usable size, same lifetime, just a different way to get there.

I suspect it just comes down to economics.  There is not enough demand for longer endurance storage to support mass production of SLC drives at a cost competitive with TLC drives, even though this should be more cost efficient where higher endurance is required.  Commercial and consumer grade customers can just buy larger drives to satisfy their longer endurance needs.  More critical customers can pay the high cost premium of true SLC drives, although some companies like Tesla have screwed this up.

Edit: Of course, that's predicated on unused space being usable for wear leveling (which apparently may not be the case) and wear leveling actually not losing data. But it's a mature technology and cheap as chips, now

A good SSD wear levels across the entire drive even if only part of the storage is in use, and should be performing scrub on read and idle time scrubbing to prevent data loss.

[aqarwaen]

i bit off topic here.but i got question let say i bought new ssd then write enough it reallocate sectors on flash.then swapped memories.
so my question is in this case would ssd/controller able to detect that i swapped memories.would it still keep or remove bad sectors if i did such thing?

[ejeffrey]

The manufacturers should provide a way to reformat TLC or QLC as MLC or SLC, for those who want to trade off capacity for more endurance.

My understanding is that MLC isn't a lot better than TLC.  So you would need to be going to pSLC to get a real benefit, and few people actually want to give up 66% of their storage.  So unfortunately we are stuck with the situation where only way to get SLC/pSLC drives is to get "industrial" SSDs that are insanely over priced and under performing.

MLC is basically in a weird untenable position for most use cases.  If you have MLC, it's generally better for performance to split the same total capacity into TLC and pSLC.  It's only where you need a specific high sustained write performance, and you don't benefit from often being able to burst faster than that.  There are applications like this, such as DVRs, data aquisition, and network capture systems.  However, they are exactly the applications where they also value high capacity.  In that case, just distributing the data across multiple TLC drives can get you the performance you need at a much lower cost/GB.

[amyk]

The manufacturers should provide a way to reformat TLC or QLC as MLC or SLC, for those who want to trade off capacity for more endurance.

They actually do, but are quite secretive about it. I just came across this:

i bit off topic here.but i got question let say i bought new ssd then write enough it reallocate sectors on flash.then swapped memories.
so my question is in this case would ssd/controller able to detect that i swapped memories.would it still keep or remove bad sectors if i did such thing?

That probably wouldn't work since they tend to store the firmware there too. But otherwise the FTL structure are stored in the flash, so if you reformat it using the manufacturer tool it'll rescan and establish new bad blocks.

[tridac]

I did that on a few older machines. One a 486dx that I keep for disk testing (suse 11.4), other utiliites and tools. The greatest gain is the bypassing disk rotational latency and seek times, both in milleseconds. Great improvement in boot times and general use, at very little expense. Otherwise, it's ex corporate server class sas 2.5 drives, which really do fly compared to 10k 2.5 types. Smart utils on all receipt, but many of the ex corporate sas drives show either zero or just a few percent wear index, and will last forever here. ZFS raidz for server and zfs mirrored for os system disks. Samsung preferably, but Micron and Toshiba sas pretty good as well....

[Geoff-AU]

Samsung .... The older 840 line of SSDs suffered a bug where cells lost chage and data that wasn't read or written often enough would vanish, and such a firmware update was needed to make the drive automatically perform rolling rewrites across itself, which increased the wear somewhat and shotened the overall lifespan accordingly as, to prevent bits getting corrupted every bit had to be rewritten every few days.

A bit hyperbolic.  The bug was that "old" data got slower and slower to read, because the drive tried many times to read the data until it succeeded.  I have an 840 with this issue, the firmware update did a full-drive refresh (which was slow, but I didn't lose any data), and the drive is still in daily use today so the overall lifespan was NOT adversely affected.  The data only has to be refreshed every 30+ days, not "every few days".  It's only a scandal because it became public.  If Samsung realised earlier that refreshing data was a thing, they would have just quietly done it and none of us would have been aware of it happening.

I wouldn't use NAND for archival storage though.  Magnetic disks are more stable (and my use-case is only for 3-5 years then I'm refreshing for bigger backup disks anyway)

[BrianHG]

The manufacturers should provide a way to reformat TLC or QLC as MLC or SLC, for those who want to trade off capacity for more endurance.

They actually do, but are quite secretive about it. I just came across this:

What a horrible esoteric hack tool and procedure used to patch the drive into SLC mode.  Talk about an error prone lengthy time pain in the ass... Though, once the drive is in SLC exclusive mode, the write performance uplift appears to function as I expected.

[David Hess]

Samsung .... The older 840 line of SSDs suffered a bug where cells lost chage and data that wasn't read or written often enough would vanish, and such a firmware update was needed to make the drive automatically perform rolling rewrites across itself, which increased the wear somewhat and shotened the overall lifespan accordingly as, to prevent bits getting corrupted every bit had to be rewritten every few days.

A bit hyperbolic.  The bug was that "old" data got slower and slower to read, because the drive tried many times to read the data until it succeeded.  I have an 840 with this issue, the firmware update did a full-drive refresh (which was slow, but I didn't lose any data), and the drive is still in daily use today so the overall lifespan was NOT adversely affected.  The data only has to be refreshed every 30+ days, not "every few days".  It's only a scandal because it became public.  If Samsung realised earlier that refreshing data was a thing, they would have just quietly done it and none of us would have been aware of it happening.

I wouldn't use NAND for archival storage though.  Magnetic disks are more stable (and my use-case is only for 3-5 years then I'm refreshing for bigger backup disks anyway)

Scrub-on-read and idle-time-scrubbing is suppose to handle that by rewriting blocks which have accumulated too many correctable errors, before they become uncorrectable errors.

I do not understand how Samsung could have been unaware of scrub-on-read and idle-time-scrubbing.

[Geoff-AU]

I do not understand how Samsung could have been unaware of scrub-on-read and idle-time-scrubbing.

I can only assume a Project Manager's deadline was threatened.

[David Hess]

I do not understand how Samsung could have been unaware of scrub-on-read and idle-time-scrubbing.

I can only assume a Project Manager's deadline was threatened.

Or they considered it a "premium" feature to be limited to data center products.  I only started using SSDs maybe 4 years ago, and at the time I evaluated various manufacturers.  Samsung discussed power loss protection only in connection with their data center products, and power loss protection is required to safely implement scrub-on-read and idle-time-scrubbing.

[tooki]

What I find unnerving is the completely misleading terminology of MLC and TLC: that suggests (and early explanations claimed) that MLC is two levels, and TLC three, which I think anyone who knows anything about electronics would take to mean voltages or charge levels or the like. That is, that excluding zero, SLC has one value, MLC two, and TLC three, giving 2, 3, and 4 possible values, respectively, including zero. But they're not. The "levels" are actually the exponent of a power of 2, so SLC has 2¹ = 2 possible values (incl. zero), MLC has 2² = 4 values, and TLC has 2³ = 8 possible values. This surely decreases S/N ratio appreciably (or whatever the correct term is in this situation; let me know!).

[David Hess]

What I find unnerving is the completely misleading terminology of MLC and TLC: that suggests (and early explanations claimed) that MLC is two levels, and TLC three, which I think anyone who knows anything about electronics would take to mean voltages or charge levels or the like. That is, that excluding zero, SLC has one value, MLC two, and TLC three, giving 2, 3, and 4 possible values, respectively, including zero. But they're not. The "levels" are actually the exponent of a power of 2, so SLC has 2¹ = 2 possible values (incl. zero), MLC has 2² = 4 values, and TLC has 2³ = 8 possible values. This surely decreases S/N ratio appreciably (or whatever the correct term is in this situation; let me know!).

You are right, and it is misleading.  I just always understood that it was a power of 2 because I do that sort of math all of the time ... to calculate things like signal to noise ratio, or how many "bits" is 40,000 counts? (15.3 bits)

2^x = 40,000
ln(2^x) = ln(40,000)
x * ln (2) - ln(40,000)
x = ln(40,000)/ln(2)
x = 15.3

Signal to noise is as good a term as any.

[amyk]

A bit hyperbolic.  The bug was that "old" data got slower and slower to read, because the drive tried many times to read the data until it succeeded.  I have an 840 with this issue, the firmware update did a full-drive refresh (which was slow, but I didn't lose any data), and the drive is still in daily use today so the overall lifespan was NOT adversely affected.  The data only has to be refreshed every 30+ days, not "every few days".  It's only a scandal because it became public.  If Samsung realised earlier that refreshing data was a thing, they would have just quietly done it and none of us would have been aware of it happening.

"in daily use" is the key word; that flash is so leaky that it can't hold readable charge for more than a month, and I bet that number would go down very quickly in a warmer environment as retention decreases exponentially with temperature.

I'm pretty sure people would've figured out sooner or later, as not everyone who has flash storage keeps it plugged in all the time. I have some old USB drives from SLC era which have gone through a lot of full-drive writes (hundreds, probably enough to wear out some modern SSDs) with no data loss after almost 2 decades. Here's a more recent case:

https://goughlui.com/2023/10/10/psa-ssds-with-ymtc-flash-prone-to-failure-check-your-ssds/

As for "TLC" "QLC" etc. it's definitely misleading, but they've used "levels" to refer to the "layers" of bits on the same page which are stored in the same cell. Programming pages happens one "level" at a time. It makes deep technical sense, but unfortunately is very useful for hiding the endurance and retention characteristics.

[tooki]

As for "TLC" "QLC" etc. it's definitely misleading, but they've used "levels" to refer to the "layers" of bits on the same page which are stored in the same cell. Programming pages happens one "level" at a time. It makes deep technical sense, but unfortunately is very useful for hiding the endurance and retention characteristics.

That sure sounds to me like it was selected very deliberately to be misleading. If it wasn’t deliberate, then it was a case of staggering incompetence.

[David Hess]

As for "TLC" "QLC" etc. it's definitely misleading, but they've used "levels" to refer to the "layers" of bits on the same page which are stored in the same cell. Programming pages happens one "level" at a time. It makes deep technical sense, but unfortunately is very useful for hiding the endurance and retention characteristics.

That sure sounds to me like it was selected very deliberately to be misleading. If it wasn’t deliberate, then it was a case of staggering incompetence.

What other terms could they have used?

Except for MLC, they unambiguously indicate how the capacity of the drive is multiplied, which is what matters to the customer.  They also give a good indication of how resilient the underlying storage is.

[tooki]

But the word “level” doesn’t normally indicate multiplication, never mind exponentiation. Parking garages often are labeled with “levels”, and if I see “parking level 3”,  I expect that to be either 2 or 3 floors off the ground (depending on whether one counts the ground level as #1). I certainly wouldn’t expect “level 3” to be 7 or 8 floors off the ground!!

I distinctly remember some early descriptions of MLC and TLC as explaining that SLC having one nonzero voltage level (let’s call the possible values 0 and 1), MLC having two nonzero voltage levels (values 0, 1, and 2), and TLC having three nonzero values (values 0, 1, 2, 3). That is clearly wrong, so clearly the naming had managed to confuse some people.

The numbering of single and triple accurately reflects bits per cell, and some name along those lines would have been more honest marketing.