Does Anyone Actually Know of a Worn Out SSD Due To Too Many Writes?

[taltamir]

i saw a guy here once who was saying his drive was running out, and it was. I suspected that lack of TRIM caused excessive write amplification which was eating up his writes. He never followed up on it for me to find out.

my own intel G2 drive is a year and a half and used used 3% of its writes.

 

[Modelworks]

Nand flash last a very long time. some hobbyist took some nand flash and repeatedly erased and set the same byte over and over until it failed. The flash was supposedly guaranteed for 100K writes. It lasted till 122K writes before failing.

 

[taltamir]

that would be 50nm SLC... 50nm MLC is 10k writes, 32nm is 5k writes, and 25nm is 3k writes.

The drive will last a long time because of wear leveling and how little the average user actually writes.

 

[IntelUser2000]

If the SSDs are like the USB thumb/jump drives, then they do have a end-of-life running period. I had used a bunch of USB drives to do folding@home for almost a year. All three of the Kingston's 4Gb drives are now totally dead. I'm questionable on going to SSD until they are more proven and the cost seems better.

Lesser known fact. They use the very low quality NAND chips for those USB flash drives. They are even ok using 3-bit MLC, while they were rejected for SSD drives. If you lasted 1 year on a USB drive, you'll probably last 20-30 years on an SSD.

 

[taltamir]

Lesser known fact. They use the very low quality NAND chips for those USB flash drives. They are even ok using 3-bit MLC, while they were rejected for SSD drives. If you lasted 1 year on a USB drive, you'll probably last 20-30 years on an SSD.

not all of them, some used high quality SLC. For a much higher price you could buy high quality SLC USB drives, which significantly outperformed their MLC counterparts.

 

[Voo]

not all of them, some used high quality SLC. For a much higher price you could buy high quality SLC USB drives, which significantly outperformed their MLC counterparts.

Sure and those quite probably use intelligent controllers that do wear leveling (uh which makes them SSDs with a different connection doesn't it?), but for the vast majority we're talking about cheap flash and cheap controller - and them dying after a year certainly enforces that theory, don't you agree?

 

[Jimmah]

Currently coming upon 1.5 years running an el-cheapo 8gb Kingston CF card as the HDD in my router. No hiccups or problems, with caching proxy set at 200mb since day 1.

My desktop drives on the other hand, both OCZ Solid 1's died at the same time, thankfully they have a 5 year warranty. Solid 2's are much better.

 

[hal2kilo]

well i've had two x25-m fail catastrophically. you might say they had accelerated death?

nasty - no SMART indicators lit up - you hit the bad sector and the controller froze. had to do the ole skip bad sector disk image - not easy these days.

As a person who has a 25x-m drive, I'm curious as to how these drives were being used, what mobo they were installed on, ect. Anecdotes have some value, but specifics would be a benefit to us all.

 

[taltamir]

Sure and those quite probably use intelligent controllers that do wear leveling (uh which makes them SSDs with a different connection doesn't it?), but for the vast majority we're talking about cheap flash and cheap controller - and them dying after a year certainly enforces that theory, don't you agree?

surprisingly no, same/similar controller with different NAND. remember you are very limited on physical space on a USB thumb drive; and its not like they have the incentive or time... And if they are using SLC which lasts at least 10x as long why would they even need it? Besides which those SLC drives existed AND stopped being made years before TRIM and full sized SSDs. I still have my 4GB SLC flash voyager GT, a few made an 8GB SLC but most companies never went above 4GB with SLC.

But I am not trying to deny that most USB flash is a cheap controller with cheap flash which are total turds that die after a year

 

[Troll Trolling]

No problems here. None whatsoever. This drive was the best $230 (in October of '09) I've ever spent on a single component.

4 Re-allocated sectors, but I'm not too worried about it. My understanding is 4 is within the acceptable range.

Gentlemen, this guy used his SSD for 16 months and wrote in it 4.27TB of data, roughly 9.1GB of data per day. Even if his flash fail at the 1000th cycle (that is very, VERY conservative); that would be 21 years of use.

Moreover, should be noted that even 25nm NAND are said to last 3000 cycles, that would make it last 3 times longer (63 years of 9.1GB/day or 21 years of 27.3GB/day).

Things get even better on the SandForce controller and on the rumored Intel's new controller (sub-one write amplification).

 

[jwilliams4200]

Gentlemen, this guy used his SSD for 16 months and wrote in it 4.27TB of data, roughly 9.1GB of data per day. Even if his flash fail at the 1000th cycle (that is very, VERY conservative); that would be 21 years of use.

The lifetime computation is not that simple. You cannot simply take the raw number of bytes he wrote and divide it by the capacity of the SSD. You also have to take into account write amplification. If all of those 4.27 TB he wrote were random 4KB writes, then the write amplification could be quite large, certainly 10 would not be a terrible guess.

Of course, he probably did not have all 4KB random writes. So the average write amplification is likely to be less than 10. Maybe less than 5. But to be conservative, you should not assume it is 1.

So 21 years is not a conservative estimate. It is likely to be half that or less.

Intel's conservative number is 20GB per day for 5 years. Still quite impressive, I think.

 

[Makaveli]

4TB's in 16 months.

I must be super conservative then i've done 1.3 TB's in 14 months.

However I have alot of tweaks done on my system.

Page file not on SSD.

Browser caches off ssd aswell.

all downloads on a storage drive.

 

[taltamir]

The lifetime computation is not that simple. You cannot simply take the raw number of bytes he wrote and divide it by the capacity of the SSD. You also have to take into account write amplification. If all of those 4.27 TB he wrote were random 4KB writes, then the write amplification could be quite large, certainly 10 would not be a terrible guess

the maximum write amplification would be 128x. This should not happen normally, but I suspect that without trim it does happen based on a guy who was here before and wore out his drive (that should have lasted longer, and did for others) in a year and a half. It requires more investigation to see if lack of TRIM really causes such a case of maxed out write amplification or not.

To be fair, he DID mention write amplification, and pointed at how low it is on some of today's top drives. But that is according to their own claimed figures, not real testing data done by impartial groups.

 

[Voo]

The lifetime computation is not that simple. You cannot simply take the raw number of bytes he wrote and divide it by the capacity of the SSD. You also have to take into account write amplification. If all of those 4.27 TB he wrote were random 4KB writes, then the write amplification could be quite large, certainly 10 would not be a terrible guess.

Actually Intel's official WA number for the standard desktop client use cases is 1.1 for their drives - I'd concede that a factor of 2 is a more reasonable guess, but no good controller should've anywhere near a WA of 10 with a reasonable amount of overprovisioning and trim.
But even with a WA of 10 we're still talking about 39TB of data you can write to a 80gb drive.

Oh and I win with 6.3TB in 17months don't I? Damn if I use the drive for another decade that way I may run out of cycles, how horrible

@Talta: Academic interest.. the Intel does still emulate 512byte blocks, so wouldn'te the worst case be 512byte and not 4kb useable data written per block? That would mean the worst case WA would be 1024 not 128 - not that it's practically relevant

 

[taltamir]

Actually Intel's official WA number for the standard desktop client use cases is 1.1 for their drives - I'd concede that a factor of 2 is a more reasonable guess, but no good controller should've anywhere near a WA of 10 with a reasonable amount of overprovisioning and trim.

intel claimed that the best a controller managed before their own came about was 40x

@Talta: Academic interest.. the Intel does still emulate 512byte blocks, so wouldn'te the worst case be 512byte and not 4kb useable data written per block? That would mean the worst case WA would be 1024 not 128 - not that it's practically relevant

You are totally right, I had not considered it and with 512B sector emulation on a 4kb sector SSD with 128 sectors min erase block (524288 bytes) you indeed get a theoretical max write amplification of 1024x... that is of course not considering the actual size of the file but its size on disk (a 10byte file takes 1 sector aka 512B on disk).

It should be noted that the TRIM thing is an educated guess, I have a working theory as to how it would cause it, but it has not been tested or confirmed (or even discussed outside of a few threads here on anandtech).
Although, IIRC intel did say they have a sub block write capability... and even without TRIM there is over provisioning to help mitigate worst case scenarios like the above. But there was this guy who was really seeing about 120 write amplification and I theorized that it was due to him not having trim and keeping the drive nearly full... I can't remember exactly what make and model he had though. I will try to look it up.

Oh and I win with 6.3TB in 17months don't I? Damn if I use the drive for another decade that way I may run out of cycles, how horrible

very nice. My G2 was bought 2009-07-23 and today (2011-02-15; 19 month later) it is:

Note that yesterday the media wear-out indicator was on 97%, dropped to 96% today after I installed a 4GB game on it. A little math shows that it has 34 more years (exactly) in it in current rate.

Anyone knows how much space is actually provisioned as spare in the intel G2 80gb? I want to calculate my write amplification.

 

[taltamir]

delete

 

[Voo]

intel claimed that the best a controller managed before their own came about was 40x

Yeah with "good controllers" I meant Intel and newer, i.e. SF and Co.

Anyone knows how much space is actually provisioned as spare in the intel G2 80gb? I want to calculate my write amplification.

Ah considering that all flash cells are sold in GB sizes, I think it's pretty definite that all manufacterers (that is except OCZ for their new drives..) just use that extra space for their consumer drives which would amount to 1024^3/1000^3-1 = 7.37%
Also confirms quite well to the always mentioned "about 8%" (considering that marketing loves to round up).


And congratulations, where do you want your throne delivered? Tztz so shameful beaten

 

[jwilliams4200]

Actually Intel's official WA number for the standard desktop client use cases is 1.1 for their drives - I'd concede that a factor of 2 is a more reasonable guess, but no good controller should've anywhere near a WA of 10 with a reasonable amount of overprovisioning and trim.

1.1 is obviously not what Intel uses to calculate the SSD lifetime. 20GB per day for 5 years comes to 3.65e13 bytes total. For an SSD with 80GiB of flash, that comes to 425.2 writes per Byte of flash. Since the flash Intel is using should have at least 5K program/erase cycles, it appears they are assuming a write amplification of 11 or 12.

 

[Voo]

1.1 is obviously not what Intel uses to calculate the SSD lifetime. 20GB per day for 5 years comes to 3.65e13 bytes total. For an SSD with 80GiB of flash, that comes to 425.2 writes per Byte of flash. Since the flash Intel is using should have at least 5K program/erase cycles, it appears they are assuming a write amplification of 11 or 12.

Actually I'm pretty sure Intel started with guaranteeing 100gb/day written data - OEMs requested 20gb/day.
Sure that was with 45nm MLC but that number should only half in that case (100gb/day sounds just like a nice round number, so who knows what their internal results were?), but I think Intel would be extremely clumsy in not including a large margin of error just to be on the save side. And heck who cares about 20gb or 100gb or 134.4gb data per day as a consumer?

No for technical details I'd believe their white papers a big deal more, or something in between.

PS: Actually I looked up the first Intel article from anand.. http://www.anandtech.com/show/2614/4
100gb/day; WA for their drives of 1.1; other competitors a WA of 10; If you want to believe some random (marketing) slides

 

[taltamir]

1.1 is obviously not what Intel uses to calculate the SSD lifetime. 20GB per day for 5 years comes to 3.65e13 bytes total. For an SSD with 80GiB of flash, that comes to 425.2 writes per Byte of flash. Since the flash Intel is using should have at least 5K program/erase cycles, it appears they are assuming a write amplification of 11 or 12.

that isn't necessarily how much they are assuming you will get, but how much they are willing to guarantee you a free replacement.
Intel warranties their CPUs for 3 years, doesn't mean they assume all CPUs will fail after 3 years... (in fact they last much longer.)

that being said, the fact that they will only warranty replace it if you gets worn out from a write amplification that is an order of magnitude larger then what they advertise is concerning. I am not buying the 1.1x figure. EDIT: ok I am buying it, seems like they were just planning for the worst. See below for my actual write amplification.

Ah considering that all flash cells are sold in GB sizes, I think it's pretty definite that all manufacterers (that is except OCZ for their new drives..) just use that extra space for their consumer drives which would amount to 1024^3/1000^3-1 = 7.37%
Also confirms quite well to the always mentioned "about 8%" (considering that marketing loves to round up).

right, then it is 5.896GB... or i can just go ahead and calculate the size. 10 chips of 8"GB" each * 1.024^3 = 85.89934592 GB
times 5000 (expected life of 34nm chips) = 429496.7296 GB life.
I used 4% (ticked over from 3% today!) of it today so that is *4/100 = 17179.869184 GB I actually wrote to it.
I wrote 8.48TB so thats 8.48*1024*1.024^3 = 9323.85860354048 GB it says I wrote to it.
My actual write amplification is: 1.8425707547169811320754716981132x
Rounds to actual write amplification = 1.84x

I DO use:
1. hibernate on SSD with hybrid sleep on and often (shouldn't matter, should be entirely sequential writes)
2. search indexing stored on SSD.
3. pagefile on the SSD.
4. browser cache on the SSD
5. thumbnails, but nothing on the SSD... all my documents that could get thumbnails are on my spindle drive or fileserver.
6. Video game save files on SSD, only select video games get installed on it

And congratulations, where do you want your throne delivered? Tztz so shameful beaten

The obvious room would be the bathroom

 

[jwilliams4200]

Actually I'm pretty sure Intel started with guaranteeing 100gb/day written data - OEMs requested 20gb/day.

1.1 is not consistent with 100GB per day. With 100GB per day for 5 years on an 80GiB SSD with 5000 program/erase cycles, the write amplification comes out to about 2.4.

So I think that 1.1 number is bogus.

 

[Voo]

right, then it is 5.896GB... or i can just go ahead and calculate the size. 10 chips of 8"GB" each * 1.024^3 = 85.89934592 GB
times 5000 (expected life of 34nm chips) = 429496.7296 GB life.
I used 4% (ticked over from 3% today!) of it today so that is *4/100 = 17179.869184 GB I actually wrote to it.
I wrote 8.48TB so thats 8.48*1024*1.024^3 = 9323.85860354048 GB it says I wrote to it.
My actual write amplification is: 1.8425707547169811320754716981132x
Rounds to actual write amplification = 1.84x

Don't see any mistake, so yeah that sounds valid. I'd think. That is as long as the TB specified by Intel are base 2 and not 10, but that sounds likely.

Btw I get a WA of somewhere 2.48 - 3.7 (only whole percentages is rather imprecise - I'll check when I get to the next whole) on my 160gb drive which is surprisingly high but then my work flow is pretty strong on small random files I think (VMs, programming stuff,..) so I'd put me on the boundary of "normal desktop useage"

All in all I think with a WA of <5 we're pretty much on the safe side of things and for most users somewhere between 1.5-2.5 seems the most reasonable guess.

 

[Catoblepas]

I saw some of you using some sort of utility to check your SSD. I have two OCZ Vertex 2s one as a boot drive and one as a faster gaming drive. I am curious would that Intel utility be able to monitor OCZ drives or would you all recommend a different utility.

 

[HeXen]

That is all true but, in practice, are there any reported cases where someone's SSB actually got to the point where worn out memory locations degraded the performance of the drive?.

how would one really know for sure without before and after benchmarks? considering the speed, it would have to be very worn to be much noticable and likely it would be gradual, so you would be gradually used to it.

sometimes i suspect my Single Cell SSD i got years ago is, but i can't prove it and i'm not even really sure honestly, i don't remember how fast it was when i got it plus i'm on a different OS now doing different things running different apps than i did back then....so how could anyone really know for sure unless it flat out dies?

 

[RU482]

I have a couple of drives showing ~30&#37; health in Crystal Disk info