Samsung 4TB QLC SSDs enter mass production

[LeetyMcLeet]

It's amazing how far SSD's have come in terms of speed and capacity. I'll probably pick one of these up to use as a games drive in next rig.

I'm still using an old (2013?) Samsung 840 Pro 250GB, that's still running perfectly and a 500GB Samsung 850 EVO for games. Reliability isn't an issue for me, as I BACK UP my data

[watercooled]

There is a reason you can buy HDD controller boards on eBay......they do go just like SSD controllers...

I've replaced a number of them over the years to recover data from a drive or even resurrect a dead drive. There is even a company in Canada who will clone a HDD controller for you for a very reasonable price, so guessing its a fairly well used service for a fairly common fault.

That's precisely the point I'm making - when HDDs do fail it's rarely instantly catastrophic. Should the need arise, replacing a controller board or even transplanting the platters is often useful as it's trivial to read the stored data. Not so with an SSD. Retrieving the mapping tables (even partial) of a fried SSD would likely be difficult at best, and you can pretty much kiss the data goodbye if not - data isn't stored verbatim on the media for a variety of reasons, the scrambling methods are non-trivial and when you add wear levelling into the mix... And that's assuming something hasn't cause the controller to instantly block erase everything, which takes a few seconds on an SSD vs many hours of a HDD (that weakness is a potential strength here).

It's amazing how far SSD's have come in terms of speed and capacity. I'll probably pick one of these up to use as a games drive in next rig.

I think people are perhaps over-estimating how much cheaper these will be. Even assuming the raw NAND price is the only factor (which it isn't) and assuming the NAND pricing will be equal per mm2 (which I doubt), you're looking at max 33% more storage for a given price. It's not exactly a huge difference, and a lot of that will likely be eaten for over-provisioning too.

I'm still using an old (2013?) Samsung 840 Pro 250GB, that's still running perfectly and a 500GB Samsung 850 EVO for games. Reliability isn't an issue for me, as I BACK UP my data

Backing up important data is sensible regardless of what media it's stored in. It's not an alternative to reliability.

[Iota]

you're looking at max 33% more storage for a given price.

I thought you could already get 2TB SSDs? So wouldn't that be 50% more storage?

[DanceswithUnix]

Exactly - although I suggest the warning with an SSD is likely to be less than with a hard drive which tend to degrade gracefully (with re-allocated sectors) rather than a sudden catastrophic failure. Of course HDDs can fail catastrophically, but I suggest it is less likely.

I think you have a rather romantic notion of HDDs


Component failure is instadeath for HDD or SSD.
Firmware bug can lead to instadeath for HDD or SSD.
Media failure (bad sector) leads to a remapped sector for HDD and SSD.

So from a top level, they seem fairly equivalent, but the devil is in the details...

SSD's used to have bad firmware bugs, which I think is still tarnishing their image even though I haven't heard of such a failure in recent drives. Early IDE drives were similar, but that is a distant memory.

On a HDD the platters are separate from the controller, so controller component failure *might* be fixed by scavenging a board from an identical drive (I've had that both work and fail). An SSD has media and controller on the same PCB, so whilst in theory you could re-solder the flash chips onto another SSD's controller to try a recovery I haven't even heard of someone even bothering to attempt it.

Sector failure on an SSD should consist of a sector failing to verify or having to lean heavily on ECC, much like a HDD, and causing a region to be remapped. On a HDD though that almost certainly means physical damage to the media which means little bits of debris being spat across the drive causing more sectors to fail later on as the bits of debris get ground across and into the platter as sectors around the reallocated sector are accessed. On an SSD, access to a block next to a worn out block has no inherent risk.

So yes a HDD can fail slowly, but it's still a failure and I find it pretty rare that I can clone the drive before complete failure as it will have "pending sectors" that are unreadable but will only be remapped when written to so just send the drive into a loop of trying to re-read the failed sectors which almost always seem to contain something important.

Storage is storage, back it up, test the backups.

I thought you could already get 2TB SSDs? So wouldn't that be 50% more storage?

This is storage per cell in the flash. So the idea is that what used to store 3TB (three bits per cell) can now store 4TB with the same amount of flash by having more voltage levels in each cell. That should drive price down, possibly make performance worse as the writing is trickier.

[Corky34]

An SSD has media and controller on the same PCB, so whilst in theory you could re-solder the flash chips onto another SSD's controller to try a recovery I haven't even heard of someone even bothering to attempt it.

I don't normally pay much attention to Linus Tech Tips as he's a bit marmite, however he did a behind the scenes of DriverSavers a while back that touches on data recovery of SSD's and replacing controllers, the whole video is worth a watch IMO but the SSD part starts around 7:35.



It's pretty amazing that they can recover data from bare flash chips and NAND flash with security chips like the iPhone.

[Ironbuket]

There is a reason you can buy HDD controller boards on eBay......they do go just like SSD controllers...

I've replaced a number of them over the years to recover data from a drive or even resurrect a dead drive. There is even a company in Canada who will clone a HDD controller for you for a very reasonable price, so guessing its a fairly well used service for a fairly common fault.

That's precisely the point I'm making - when HDDs do fail it's rarely instantly catastrophic. Should the need arise, replacing a controller board or even transplanting the platters is often useful as it's trivial to read the stored data. Not so with an SSD. Retrieving the mapping tables (even partial) of a fried SSD would likely be difficult at best, and you can pretty much kiss the data goodbye if not - data isn't stored verbatim on the media for a variety of reasons, the scrambling methods are non-trivial and when you add wear levelling into the mix... And that's assuming something hasn't cause the controller to instantly block erase everything, which takes a few seconds on an SSD vs many hours of a HDD (that weakness is a potential strength here).

I thought it was a myth that you can simply switch controller boards on physical hard drives and it will work straight away. Dont you have swap over the bios chips or reprogram them before they will work?

[watercooled]

I thought you could already get 2TB SSDs? So wouldn't that be 50% more storage?

I said 'for a given price' - not max possible capacity which is also dependent on what the controller supports i.e. number of NAND channels and number of dies per channel. But all else being the same you'd still get max 33% more total capacity for the same cell count (though they're likely using different dies for QLC anyway). As DanceswithUnix says it's about how much data you can fit per NAND cell - QLC gets you just 33% more raw capacity (likely even less when you add more robust ECC, over-provisioning, etc.)

SLC>MLC you get double (100% more)
MLC>TLC you get 50% more
TLC>QLC you get 33% more

And despite the diminishing returns in terms of capacity, the number of possible states and therefore complexity continues to double:
SLC has 2 possible states
MLC has 4 possible states
TLC has 8 possible states
QLC has 16 possible states

And the increased complexity of reading the states at increasing precision also has an impact on endurance, which is AFAIK actually worse than half each time.

I thought it was a myth that you can simply switch controller boards on physical hard drives and it will work straight away. Dont you have swap over the bios chips or reprogram them before they will work?

It depends on the failure mode and whether you can get a compatible board (it's sometimes not just the same retail model number you need to check). I've seen it work enough times for it to not be a 'myth'.

[watercooled]

Component failure is instadeath for HDD or SSD.
Firmware bug can lead to instadeath for HDD or SSD.
Media failure (bad sector) leads to a remapped sector for HDD and SSD.

So from a top level, they seem fairly equivalent, but the devil is in the details...

I don't agree that component or firmware errors = instadeath for a HDD. Perhaps the usability of the drive but not necessarily the data stored within. A controller board or firmware failure doesn't instantly destroy what's on the platters, recovery of which is generally fairly straightforward with the right equipment. SSD firmware/controller bugs can and have caused instadeath for user data. The fact SSDs can secure erase in a matter of seconds is proof that this failure mode is possible, however unlikely with more carefully designed firmware etc. And don't forget not all SSDs are produced by the big, experienced companies - compared to HDD production, assembling SSDs is relatively straightforward allowing many smaller companies to give it a go. OTOH HDD platters can shatter if subject to enough force, though again that's pretty unlikely unless you throw a bare drive on to a concrete floor or something.

Don't get me wrong I'm not suggesting they're inherently less reliable than HDDs for storage given they're not vulnerable to a number of HDD failure modes, but it's still not something to simply ignore. Both have inherent advantages and disadvantages. It's interesting from an academic point if nothing else. Some people will outright refuse to accept or discuss any drawbacks of SSDs though, however small.

If everyone kept minute-by-minute backups then data recovery firms wouldn't be as big of an industry as they are.

As an aside, I know they're unfortunately not produced to the same standards as SSDs but some USB drives have a tendency to just outright stop working with absolutely no warning. IMO companies/controller manufacturers would do well to aim for less catastrophic failure modes rather than just assuming they're used as disposable devices with no important data carried on them.

[aidanjt]

No storage medium or RAID or no RAID is a substitute for a proper backup strategy.

[azrael-]

I said 'for a given price' - not max possible capacity which is also dependent on what the controller supports i.e. number of NAND channels and number of dies per channel. But all else being the same you'd still get max 33% more total capacity for the same cell count (though they're likely using different dies for QLC anyway). As DanceswithUnix says it's about how much data you can fit per NAND cell - QLC gets you just 33% more raw capacity (likely even less when you add more robust ECC, over-provisioning, etc.)

SLC>MLC you get double (100% more)
MLC>TLC you get 50% more
TLC>QLC you get 33% more

And despite the diminishing returns in terms of capacity, the number of possible states and therefore complexity continues to double:
SLC has 2 possible states
MLC has 4 possible states
TLC has 8 possible states
QLC has 16 possible states

And the increased complexity of reading the states at increasing precision also has an impact on endurance, which is AFAIK actually worse than half each time.


It depends on the failure mode and whether you can get a compatible board (it's sometimes not just the same retail model number you need to check). I've seen it work enough times for it to not be a 'myth'.

Was just going to bring up the bit about 2^4 (16) possible states with QLC. Those differences in voltage are extremely difficult to handle so tbh with QLC I'm as much afraid of the low-ish endurance as I am of data deterioration. It's quite telling that warranty is significantly lower than with TLC drives.

[DanceswithUnix]

I don't agree that component or firmware errors = instadeath for a HDD

Early IDE drives stored part of their firmware on one of the platters, and would occasionally erase it when sent odd commands bricking the drive, which Windows 3.1 of the era was prone to doing.
The "IBM Deathstar" problem was a firmware bug to do with how the drive idled.

Modern SATA drives don't do that, so it would be unfair to say hdds are bug ridden junk. I just think the same goes for SSDs, the blight of SF2281 based drives in 2011 shouldn't be the standard by which modern drives are tarred but all too often they are.

and yes I've seen hard drives with big scorch marks on the PCB and components cracked in half from the heat of their failure brought back with a simple controller board swap, but that is rare compared to the platters failing and the creeping death of pending/remapped sector counts climbing.

The fact SSDs can secure erase in a matter of seconds is proof that this failure mode is possible, however unlikely with more carefully designed firmware etc.

And hard drives can't do the same?

http://lmgtfy.com/?q=hdd+instant+secure+erase

OTOH HDD platters can shatter if subject to enough force,

Like the company who allowed someone to let rip with a pneumatic drill next to a large raid array?

And yes USB drives are bad, but I find micro SD cards are awful, even the top branded ones bought from reliable channels so you know they aren't cheap knock-offs. I have enough Raspberry Pi like devices that boot off them that I see card failures far too often. But like modern consumer hard drives, they are made down to a price not up to a standard.

[peterb]

And hard drives can't do the same?

http://lmgtfy.com/?q=hdd+instant+secure+erase

I think the point was that overwriting a hard drive takes a long time

try

dd if=/dev/zero of= /dev/sda bs=4096

on a 4TB HDD.

while an SSD is much quicker (by the nature of the device)

So there is the possibility of getting some data off the HDD if stopped quickly enough.

Whether the erase command is likely to be issued 'by mistake' in real life is a moot point though.

[CAT-THE-FIFTH]

I would like to see more tests regarding data retention on modern SSDs.

[peterb]

I would like to see more tests regarding data retention on modern SSDs.

Yes, its not something that features in specifications.

I was looking at clonning an OS disk for backup (so if the OS disk fails, I can swap in a replacement ready to go. Initial thought were that an SSD would be ideal - robust, light weight etc - but data retention is a concern - cheaper and probably more reliable long term to use two or three hard drives.

The data isn't critical - the OS can always be reloaded - but have a clone ready to go saves a lot of faffing about when the installed drive fails (as it will)

[DanceswithUnix]

while an SSD is much quicker (by the nature of the device)

That's still arguing that firmware errors are worth worrying about, and they just aren't. Not that drives are bug free, they are riddled with them like any modern device but the emphases will always be on keeping user data safe so those bugs are rare.

If you want to catch accidental "dd" style erasing of a device, a modern hdd can sequential write at over 200MB/s as they are really good at that, so about half the speed of a SATA SSD. You have to be similarly fast on the power chord in both cases, and will almost certainly be down to backups.


Firmware is an interesting thing though, I can't remember the last time I upgraded the firmware on a spinning disk (probably an IBM for the Deathstar problem). Every SSD in the house has had at least one update though. I mean, I just don't even look for updated firmware for hard drives, but I do for just about everything else, is that just me?

[DanceswithUnix]

I would like to see more tests regarding data retention on modern SSDs.

I don't think you will, it comes down to extended life tests and trusting the manufacturer data as by the time an independent reviewer has spent 3 years testing a product it will be obsolete an no-one will care about the results.

https://www.anandtech.com/show/9248/...data-retention

Initial thought were that an SSD would be ideal - robust, light weight etc - but data retention is a concern - cheaper and probably more reliable long term to use two or three hard drives.

See the above, it is all about temperature. Storing in a greenhouse is bad, but if you were after an excuse for that wine cellar you always wanted to get built...
Edit: Though hard drives are cheaper, and you can spend the difference on wine