Disturbing trends in NAND flash...

[amyk]

I was looking around at information on SSD reliability and stumbled across this:

http://www.jedec.org/sites/default/files/Alvin_Cox%20%5BCompatibility%20Mode%5D_0.pdf

I know NAND flash naturally erases itself over time from electrons escaping the floating gates, but slides 56 and 57 that show the actual rates are a bit disturbing; apparently "client" class SSDs are only expected to retain data for 1 year turned off? And the best retention is achieved by heating the flash when it's in use, then cooling it down when powered off? The worst thing is, modern SSDs store the bulk of their firmware in the main flash, so even if you keep a blank SSD with no other data on it, it may eventually become unusable since the firmware has erased itself!

Not too long ago I read in some of Samsung's datasheets that they specified retention for 5 and 10 years; this isn't forever, but still much more reasonable than 1 year or 3 months. I have magnetic media from the 80s that still retains its data, and know that magnetic and optical media typically lasts for decades if well-kept, but the longevity of flash scares me. I have SD cards and other flash-based storage that I regularly back up to magnetic hard drives. I thought consumers preferred storage devices that would actually keep their data for many years, but apparently capacity (and speed) is everything these days, at the expense of longevity...

[dfmischler]

This certainly answers the questions about whether you should try to keep long term backups or archives on flash devices, huh?

[Fsck]

well people want smaller and smaller geometries and more and more states per cell.

but, I'm pretty sure those are just minimum specifications.

[angst7]

I think the main takeaway from the data retention slide is that you should avoid leaving your SSD in your car in summertime.

[peter.mitchell]

The worst thing is, modern SSDs store the bulk of their firmware in the main flash, so even if you keep a blank SSD with no other data on it, it may eventually become unusable since the firmware has erased itself!

Are you sure this is the case? It seems inherently dangerous. Eg, the controller does all of the error correction and management, but the controller's software is stored in the storage it does error correction on...
Additionally, I am under the impression that reads don't refresh standard nand flash... not particularly sure what this implies.

But yes, it is true, for a long term storage flash is undesirable, there are a number of reasons aside from just time durability though.

[rollatorwieltje]

I have thrown away plenty of floppies and CDs that became unreadable. Also VHS and cassette tapes where the quality has degraded so badly that they were practically unusable. Only Hi8 tapes seem to last forever without significant quality loss, but try to find something that plays them these days (especially with the shortplay/longplay crap that was going on).

I don't mind it really. I would buy a bigger/faster SSD over one that guarantees 5 year data retention. I have a file server for important stuff and change the harddrive every 2 years. Just consumer grade stuff, not even server drives. So far only one drive has failed in about 10 years, but was still recoverable (bad bearing, with a gentle tap it still worked).

[Lurch]

Only Hi8 tapes seem to last forever without significant quality loss, but try to find something that plays them these days (especially with the shortplay/longplay crap that was going on).

I have some Hi8 tapes. I also have a Sony Hi8 Video Walkman thingy.

If you are going to pick a medium to use as an archive then make sure you can read it and have a contingency plan!

[dexters_lab]

while i am aware of the endurance limitations of nand i wasn't aware of the temperature issue... there is no reference to cell size in that document i wonder if it is a factor

could this problem be less pronounced on older, less dense nand?

either way, it's time nand was laid out to pasture!

[amyk]

could this problem be less pronounced on older, less dense nand?

Yes, endurance and retention were much higher for flash in bigger process sizes, and partially for SLC instead of MLC. E.g. the ancient K9F3208W0A from 2000 (probably >100nm process), which is only 4MB but has 10 years retention and 1 million endurance.

There are also low-cost grades of flash such as K9L8G08U0M which is 5 years retention, but only 1.5K endurance according to the datasheet. This is 90nm 2-level MLC from 2005, so I really wonder what the latest low-cost 21nm 8-level TLC now has... 1 year retention and <1K endurance? Hard to find info as the manufacturers obviously don't feel like saying... but this article suggests that a few hundred cycles is about right.

[peter.mitchell]

i thought write endurance for latest flash was something around 80k....

[Psi]

i thought write endurance for latest flash was something around 80k....

"latest flash" doesn't necessarily mean better flash.

There are factory's producing new generation flash intended to be cheap for cheap products. ie crap.

[A Hellene]

As very rightly stated at the Bunnie's blog presentation about the :

-- Faking Reliability:
Flash memory is incredibly unreliable.
- You are not storing data, you are storing probabilistic approximation of your data.
- Workaround: Computational Error Correction (ECC).

Read this: Understanding TLC NAND: This was the reason why I chose to replace my 'tired' Crucial m4 (MLC) SSD (due to the >2800 erase cycles per block average out of the max. of 3000 cycles, after 3+ years of 24/7 operation as a system disk at my workstation) with an MLC SSD instead of its slightly faster TLC counterpart.


-George

[casper.bang]

I was looking around at information on SSD reliability and stumbled across this:

http://www.jedec.org/sites/default/files/Alvin_Cox%20%5BCompatibility%20Mode%5D_0.pdf

I know NAND flash naturally erases itself over time from electrons escaping the floating gates, but slides 56 and 57 that show the actual rates are a bit disturbing; apparently "client" class SSDs are only expected to retain data for 1 year turned off? And the best retention is achieved by heating the flash when it's in use, then cooling it down when powered off? The worst thing is, modern SSDs store the bulk of their firmware in the main flash, so even if you keep a blank SSD with no other data on it, it may eventually become unusable since the firmware has erased itself!

Not too long ago I read in some of Samsung's datasheets that they specified retention for 5 and 10 years; this isn't forever, but still much more reasonable than 1 year or 3 months. I have magnetic media from the 80s that still retains its data, and know that magnetic and optical media typically lasts for decades if well-kept, but the longevity of flash scares me. I have SD cards and other flash-based storage that I regularly back up to magnetic hard drives. I thought consumers preferred storage devices that would actually keep their data for many years, but apparently capacity (and speed) is everything these days, at the expense of longevity...

OTOH I have had about 7  SSD's (top tier brands like Intel and Samsung) in the last 5-6 years abd never had a problem. I've lost 2 magnetic drives in the past where the reading head would park itself into the spinning platter, so not sure I am quite as skeptical. Even Samsungs latest TLC versions seems to do well. I have had 10 consumer MLC disks on a high-workload server (40-50GB rewritten a day) for a year with degradation now down to 82%, unlike old magnetic storage, you get some nice predictability with SMART monitoring. SD cards abd USB sticks are an entirely different matter of course!

[casper.bang]

As very rightly stated at the Bunnie's blog presentation about the :

-- Faking Reliability:
Flash memory is incredibly unreliable.
- You are not storing data, you are storing probabilistic approximation of your data.
- Workaround: Computational Error Correction (ECC).

Read this: Understanding TLC NAND: This was the reason why I chose to replace my 'tired' Crucial m4 (MLC) SSD (due to the >2800 erase cycles per block average out of the max. of 3000 cycles, after 3+ years of 24/7 operation as a system disk at my workstation) with an MLC SSD instead of its slightly faster TLC counterpart.

Fair enough (I would also pay slightly extra for MLC over TLC) but have you seen the amount of error correction on an old spinning rust disk? Try to fire up Spinrite or a similar tool, it too is massively scary to watch!

[A Hellene]

I agree with that, Casper! This is why I have chosen to use as data HDDs only high quality ones.

For the record, my workstation Win7-based PC has four terabyte-class HDDs (of the WD Black family drives, with five years of warranty) as data and backup disk drives, and the SSD is used only for the OS and the programs, without a swap file (which exists emulated in RAM for the programs that might need it).

Truth is that since I begun to keep the computer always on (meaning running 24/7 and on a beefy UPS) since 2002 or something like that, and setting it to be falling to sleep mode a few minutes after user inactivity instead of booting it when I needed it, I did never loose another HDD! Not a single one! Not ever, since then! Waking the system up at any time is a matter of a couple of seconds delay time, with the monitor (a DELL U2413) being the laziest peripheral to wake up.


-George

[amyk]

Fair enough (I would also pay slightly extra for MLC over TLC) but have you seen the amount of error correction on an old spinning rust disk? Try to fire up Spinrite or a similar tool, it too is massively scary to watch!

I would pay extra for SLC over MLC, since I think the 10x endurance and retention over MLC is well worth the 2-3x price/capacity increase. (Or if the latest articles I could find are to be believed, which states that SLC still has 100K P/E cycles... that would be a 100x endurance ) Speed and capacity matters but to me longevity matters even more. It gets more and more difficult to find older NAND in the medium capacity range - 1-16Gbits - while flash memory at the low end (kilobits to megabits) are still plentiful, serial flash/EEPROM with endurance and longevity that is much higher. It's just a little sad that we seem to be heading in the direction of bigger and bigger flash memories with shorter and shorter lifetimes. The data erases itself, so you have to keep it powered to periodically rewrite the data, but then rewriting wears out the flash - looks like a horrible way to store something, if you ask me.

At the moment the closest alternative to flash for nonvolatile, random-access shock-resistant storage seems to be F-RAM but its density is still far too low - 8MBit is the largest I could find.

[Jon86]

Doesn't worry me much to be honest, my SSD is for my OS and programs that I want to run quickly.
If I had data that was actually important, I'd have several drives in RAID anyway.

Anyway, I highly doubt you'd lose anything after a year or so of having your SSD turned off.

[Rufus]

I think the 10x endurance and retention over MLC is well worth the 2-3x price/capacity increase.

I have a 160G Intel SSD http://www.tomshardware.com/reviews/Intel-x25-m-SSD,2012.html which has been used on my main PC running Win7 as a system disk, and scratch disk, with swap file and temp directory for about 4 years. It has racked up about 1500 power cycles and 25,000 power on hours. It reports 1 reallocated sector and 98% life remaining.

For typical PC usage endurance isn't close to being an issue.

[Zero999]

Why do people put swap files on SSDs? It sounds silly. If you keep going into swap, buy more RAM, rather than wearing out your drive.

[Rufus]

Why do people put swap files on SSDs? It sounds silly. If you keep going into swap, buy more RAM, rather than wearing out your drive.

Because Windows (rightly) almost insists on having one. Paging out rarely used memory frees it for better use without having to 'keep' going into swap. Under exceptional circumstances the machine slows down rather than crashes due to lack of memory. Programs can reserve large amounts of memory which never gets accessed, reserving RAM to never be used is much more of a performance hit (or expensive) than reserving swap file space.

I mentioned swap file because some people foolishly do not place it on SSDs thinking the wear it causes will be significant.

[rolycat]

At the moment the closest alternative to flash for nonvolatile, random-access shock-resistant storage seems to be F-RAM but its density is still far too low - 8MBit is the largest I could find.

If you are awaiting a more reliable alternative to flash, how about phase change memory (PCM)? Micron have a 1Gb PCM chip (co-packaged with 512Mb of DRAM) in production, and it has been used in a Nokia mobile: Link

A little further off, the long-awaited MRAM ''universal memory' seems destined to finally hit the mass market in a few years, if this report is accurate.

[SeanB]

I tried a while ago using a somewhat used SD card as an OS install, and created a swap partition on it. Installed and left running, and after about 3 days the SD card had stopped responding, giving me a nice kernel panic to a black screen. Figured that the EEPC was not going to run Ubuntu reliably, so switched to Puppy Linux, running off a SD card ( and when booted I can remove the SD card and nothing happens other than I can use the SD reader then) and it runs fine without a swap partition.

The SD card was fine after a reboot, though the swap was unreadable.

[XOIIO]

A little further off, the long-awaited MRAM ''universal memory' seems destined to finally hit the mass market in a few years, if this report is accurate.

I can't wait until we have pieces of glass that store insane amount of data like in minority report.


Anyways, I'm surprised that some flash has such a short lifespan on the data it holds, I'm sure that hard drive companies wouldn't use that since it is possible a drive in stock could just knock itself out, though I guess some companies could use this as a possible way to mess with people's warranties, who knows. I doubt this will affect the rise of popularity in SSD's though, I doubt many computer owners who would invest in one would have their computer turned off for an entire year.

[casper.bang]

Why do people put swap files on SSDs? It sounds silly. If you keep going into swap, buy more RAM, rather than wearing out your drive.

Now you are being silly. Waking up from hibernation (default when you close the lid of a Mac or Linux box) would be stupidly slow with a spinning rust disk.

[dexters_lab]

I tried a while ago using a somewhat used SD card as an OS install, and created a swap partition on it. Installed and left running, and after about 3 days the SD card had stopped responding, giving me a nice kernel panic to a black screen. Figured that the EEPC was not going to run Ubuntu reliably, so switched to Puppy Linux, running off a SD card ( and when booted I can remove the SD card and nothing happens other than I can use the SD reader then) and it runs fine without a swap partition.

The SD card was fine after a reboot, though the swap was unreadable.

SD cards would have less spare blocks for remapping and probably simpler wear levelling than a real SSD

i ran windows 2003 server edition as a webserver from a 16gb compact flash, it ran fine for several weeks but i got the jitters after looking at the almost continual barrage of disk writes so swapped it for a 2.5" disk before it died.

my desktop and laptop both have 240gb OCZ SSDs, both have the swap file turned off as i have plenty of ram for my needs. Another important factor is not to fill up an SSD, the more unused blocks the better the wear levelling will work