I Read the summary. At the end of it was an ellipsis, so 'read more' show follow on from it, right? But then the article page doesn't continue from it, they don't even share the same first paragraph!?
We made a choice a month or so back to condense "news" stories into a smaller format on the main page, at least in the combined view. If you look above the stories you'll see buttons for "Just Reviews" and "Just News". Click "Just News" and you'll see the full summary. I like the current design, mostly because as we get more news stories they won't overwhelm the major articles. Cheers!
Maybe the button should say "Full Article", instead. I like the synopsis approach for the front page, but the differing content *does* lend to a feeling on non-continuity, since one generally expects the type of setup the OP alludes to.
Anand, you do a good job reporting on the SSD market. Solid state storage makes a pretty good subject for more in-depth reporting. I'm eagerly awaiting the next crop of drives to get to your bench.
Yes, this is what I've been waiting for... not necessarily a rough release date or anything but just a detailed update about the 25nm Intel drives.
When SSD's were just starting to get publicity, I was a skeptic. However, I was aiming to build a monster of a PC around that time and decided to get an Intel X-25M (which I thought would be much more of a niche product). It was definitely one of the best purchases I made however, mostly because I NOTICED the speed and responsiveness in everything from everyday tasks like opening windows to gaming load times.
Anyways, just wanted to say "Thank You!" for your in depth diagrams and amazing reviews. (even though this is just a brief one) When I read your articles, I feel not only that you know a lot about what you're talking about, but also that you ENJOY this and get as excited as any PC enthusiast would.
Isnt HP working in partnership with Micron making Memresistor based storage. How close is that to actual production samples? If Memresistors are cost feasible then wouldnt NAND based storage become a no go. I wonder if Anandtech has heard anything about this technology.
Thanks for this article, interesting stuff! Keep working on such smaller "Tech"-articles regarding SSDs, they are really interesting and allow us a view under the hoods of SSDs.
I didn't realize that NAND P/E cycles were dropping so precipitously. No amount of error correction is going to help you when you start running out of P/E cycles. Suddenly, I don't feel so bad about only having a first-gen Intel SSD; apparently, the new third gen drives will have less than a third the lifespan at similar capacity sizes.
Doesn't reducing this to 3000 P/E cycles significantly reduce the usefulness of these drives? If memory serves, Anand mentioned in the original Intel SSD launch article that Intel intended for the drives to survive 100GB of writing a day for 10 years. On an 80GB drive, that would require at least 4562 P/E cycles, which was perfectly doable when the drive had 10000 P/E cycles... It seems like Intel has abandoned this goal on anything smaller than a 160GB drive.
It would also depend on the write amplification of the controller. If Intel can get the amplification of the gen 3 controllers down, similar to what SandForce is doing, then it might be possible for low amplification gen 3 drives to last just as long as gen 1 drives.
It does make you wonder though.I'd be interested to see what starts failing first, SF drives or Intel ones. Have there been any long term studies on this sort of thing?
Out of curiosity, the ECC is a 24-bit BCH code, so what's the rate of the code? It's a rate k/24 code presumably, so what is k? Are there any more details on this?
What with the rated P/E cycles dropping so quickly with the size of the transistors, it may be interesting to see how complicated the ECC will need to become. But first we need a reference point on what they are doing now.
Thanks for a great article as usual. I have a few questions that I hope you (or anyone else) enlighten me on.
1. Is a 3,000 P/E cycle rating an average per cell? Is there typically high variation, with some cells lasting 4,000 cycles and some lasting 2,000, for example, or do they all fail after roughly the same number of cycles?
2. When cells run out of cycles, what happens? Does the controller always recognize the bad cells and start reducing drive capacity or spare area, or is there a chance of losing data when it thinks it's written to an unwriteable cell?
3. On a "typical" 34 nm SSD (like Intel X-25M) what's the likelihood of a soft error changing a single bit in my data? Is it something like 1 error per X GB or is it practically zero until NAND degrades after enough use? If I understand correctly, the RAISE error correction graph implies it's a function of time, and the table below the graph shows the probability of uncorrectable errors hitting the end user is significant.
Finally, a suggestion. Data integrity is important to a lot of SSD users (even consumers), and I know you test reliability as much as possible and have shown older SSDs getting bricked or becoming unuseable. Many of us do nightly backups and aren't that scared of total failure. I'm more worried about data being slowly corrupted as a bit changes here or there. Would it be difficult to add to your testing suite the equivalent of memtest for SSDs? It could write and rewrite patterns to the drive constantly for a week or so and see if anything out of the ordinary comes up.
With all the competition I think SSD makers have a incentives to sacrifice "soft error" reliability in favor of price and capacity, and that worries me.
Variation is suppsed to be very high, with some producers selling "Select" SLC flash models, where 93% of cells can withstand 2 million cycles. It goes into enterprise drives like STEC, while the non-select, but still within spec one one goes to fast Compact Flash or SD cards.
Endurance is not the only parameter affected by going analog in MLC. Also suceptibility to electric noise, that may cause misread of a charge level. This is also getting worse with higher densities. Three-bit cells are even worse here.
In enterprise systems when the cells of an SSD or a flash module are near the wear limit the system log of your array informs about this, or even e-mails message about this. Even if it's ignored then there's always redundancy of RAID - SSD will be simply swapped to another, new SSD, that begins wear-process from the start.
With the last generation of flash, manufacturers stopped talking about SLC and started pushing Enterprise MLC. Is there some reason SLC is loosing out? I would think that at smaller nodes, SLC would not only become affordable, but might actually start to make sense with it's significantly higher endurance.
Yes, SLC is losing out because of the cost. SLC drives are still available, but they are generally only for extreme usage cases where the usage justifies the cost.
A hard drive in your laptop is written to far less than the capability of MLC flash -- so there's no reason to double the cost for endurance that will never required.
Enterprise MLC is a nice solution for those who demand beyond normal MLC capabilities but not full-fledged SLC pricing.
Longer term, I think you'll continue to see diversification in the market. Perhaps Three-level cell SSDs for consumers who are extremely price sensitive vs higher endurance solutions for those willing to pay more.
It seems to me that no matter how sophisticated controllers and memory becomes, the decreasing writability of the cells will overcome all of this. We're dependent on those numbers, and at some point the write cycle max will be reached. I don't quite understand how a controller can get around it. All they can do is to use the cells more efficiently. At some point, there will be no more gains possible, and going to smaller cells will result in decreased lifetime, requiring so much spare capacity on the drive that the point of using smaller process tech will have been negated. Sandforce already uses up to 28% spares on it's better drives. Where will that go at 25nm? What about the less expensive models such as the OWC with 7% I just bought?
If they can't get 25nm much better than 3,000 writes, then the next smaller node may be no better than about 1,500. Can they really deal with that? I don't see how.
Let's assume 100MB/sec is an attainable sustained rate, and generate/write randomized data repeatedly over the whole disk as a single 64GB file and then delete it, and do it again. 64000 / 100 = about 11 minutes for one cycle. It would take a little over 3 weeks to hit a 3,000 limit.
Sounds bad. But, my laptop disk has written about 64GB in *2 weeks*: 1800 times slower usage. If it was an SSD it would last a hundred years at that rate.
There are some software problems involved. Windows frequently moves a lot of files through main disk drive even if it doesn't need it.. For example when using decompression using copy-and-paste in some decompression programs the files are first decompressed to %TEMP%, and only then moved to final location. This doesn't happen if you use "Extract" from program menu, it's the issue with Windows Explorer.
Hey Anand, Are you able to test endurance? You are not, correct? Because the wear leveling algorithm is too good. You cannot R/W the same 1MB of data 10,000 times because the wear leveling algorithm will spread that wear across the entire drive. You could try randomizing the data as it is written (because the SandForce controller will cheat and say data has been written when it has not actually been if the data is the same or even similar). Even if you create a small partition and write to it over and over, that still will not work. Correct?
Is there a way you can write to the drive in raw mode, testing endurance that way?
One can certainly purchase raw Nand and use a Flash programmer/reader to cycle a small portion of the Nand device and measure Raw Bit Error Rates as a function of cycle.
Obviously those in the industry do this all the time.
Interesting, if higher yields are causing more errors, I might wanna buy Intel 32nm processors NOW before the 25 nm come out with higher error rates, and lower clock speeds!
About the drives, 32nm is already pretty awesome I'd say! I still own a 72nm ssd somewhere!
Didn't your (or intel's) slide state that 3rd gen ssd would increase life span from 7.5TB - 15TB to 30TB - 60TB. Both have 80gb model as smallest so smallest ssd drives lifespan should go way up (4x).
" today 25nm MLC NAND is good for around 3,000 program/erase cycles per cell.:"
I see a lot of statistics on SSD regarding life of an SSD, and I consider myself in the extremely large mass of people still watching the SSD market to find when the value mark hits and I will take the plunge into SSD.
For the vast majority of us, SSD still don't make sense based on $ and increased performance. (Note I fully expect the ANAND readership to disagree on their personal experiences with SSD and how great they are at improving performance of their machine). However plunking down another 25% in a system build and getting 1/4 the size of a standard drive and still having to buy a standard drive as a second in the system just seems like too much at the pace the industry is changing. I mean with my last computer lasting 5 - 7 years (I know there will be those of you who groan at this) I still consider this an average life cycle for most people. That said If I plunk down this now, I need to know that it will last my 5 - 7 years. I don't want to have to re-invest in another SSD in 3 years because of decreased read/write cycles on each new generation.
All this leads to my question: What do these numbers mean to an average person like me that plays games for an hour or two every night, and uses the web for email and news???
To make an informed decision on SSD it's imparitive that we can weigh our own use versus product specs, but these numbers just are too big to make sense of... maybe there is a rule of thumb to use?
We’ve updated our terms. By continuing to use the site and/or by logging into your account, you agree to the Site’s updated Terms of Use and Privacy Policy.
Facebook
Twitter
RSS
24 Comments
Back to Article
blyndy - Thursday, December 2, 2010 - link
I Read the summary. At the end of it was an ellipsis, so 'read more' show follow on from it, right? But then the article page doesn't continue from it, they don't even share the same first paragraph!?JarredWalton - Thursday, December 2, 2010 - link
We made a choice a month or so back to condense "news" stories into a smaller format on the main page, at least in the combined view. If you look above the stories you'll see buttons for "Just Reviews" and "Just News". Click "Just News" and you'll see the full summary. I like the current design, mostly because as we get more news stories they won't overwhelm the major articles. Cheers!ggathagan - Thursday, December 2, 2010 - link
Maybe the button should say "Full Article", instead.I like the synopsis approach for the front page, but the differing content *does* lend to a feeling on non-continuity, since one generally expects the type of setup the OP alludes to.
ckryan - Thursday, December 2, 2010 - link
Anand, you do a good job reporting on the SSD market. Solid state storage makes a pretty good subject for more in-depth reporting. I'm eagerly awaiting the next crop of drives to get to your bench.Right on.
redechelon - Thursday, December 2, 2010 - link
Yes, this is what I've been waiting for... not necessarily a rough release date or anything but just a detailed update about the 25nm Intel drives.When SSD's were just starting to get publicity, I was a skeptic. However, I was aiming to build a monster of a PC around that time and decided to get an Intel X-25M (which I thought would be much more of a niche product). It was definitely one of the best purchases I made however, mostly because I NOTICED the speed and responsiveness in everything from everyday tasks like opening windows to gaming load times.
Anyways, just wanted to say "Thank You!" for your in depth diagrams and amazing reviews. (even though this is just a brief one) When I read your articles, I feel not only that you know a lot about what you're talking about, but also that you ENJOY this and get as excited as any PC enthusiast would.
arnavvdesai - Thursday, December 2, 2010 - link
Isnt HP working in partnership with Micron making Memresistor based storage. How close is that to actual production samples? If Memresistors are cost feasible then wouldnt NAND based storage become a no go.I wonder if Anandtech has heard anything about this technology.
DanNeely - Thursday, December 2, 2010 - link
Apparently not until 2013http://www.technologyreview.com/computing/25018/?a...
Chloiber - Thursday, December 2, 2010 - link
Thanks for this article, interesting stuff!Keep working on such smaller "Tech"-articles regarding SSDs, they are really interesting and allow us a view under the hoods of SSDs.
Guspaz - Thursday, December 2, 2010 - link
I didn't realize that NAND P/E cycles were dropping so precipitously. No amount of error correction is going to help you when you start running out of P/E cycles. Suddenly, I don't feel so bad about only having a first-gen Intel SSD; apparently, the new third gen drives will have less than a third the lifespan at similar capacity sizes.Doesn't reducing this to 3000 P/E cycles significantly reduce the usefulness of these drives? If memory serves, Anand mentioned in the original Intel SSD launch article that Intel intended for the drives to survive 100GB of writing a day for 10 years. On an 80GB drive, that would require at least 4562 P/E cycles, which was perfectly doable when the drive had 10000 P/E cycles... It seems like Intel has abandoned this goal on anything smaller than a 160GB drive.
Mr Perfect - Thursday, December 2, 2010 - link
It would also depend on the write amplification of the controller. If Intel can get the amplification of the gen 3 controllers down, similar to what SandForce is doing, then it might be possible for low amplification gen 3 drives to last just as long as gen 1 drives.It does make you wonder though.I'd be interested to see what starts failing first, SF drives or Intel ones. Have there been any long term studies on this sort of thing?
michaelaj - Thursday, December 2, 2010 - link
Out of curiosity, the ECC is a 24-bit BCH code, so what's the rate of the code? It's a rate k/24 code presumably, so what is k? Are there any more details on this?What with the rated P/E cycles dropping so quickly with the size of the transistors, it may be interesting to see how complicated the ECC will need to become. But first we need a reference point on what they are doing now.
campbbri - Thursday, December 2, 2010 - link
Anand,Thanks for a great article as usual. I have a few questions that I hope you (or anyone else) enlighten me on.
1. Is a 3,000 P/E cycle rating an average per cell? Is there typically high variation, with some cells lasting 4,000 cycles and some lasting 2,000, for example, or do they all fail after roughly the same number of cycles?
2. When cells run out of cycles, what happens? Does the controller always recognize the bad cells and start reducing drive capacity or spare area, or is there a chance of losing data when it thinks it's written to an unwriteable cell?
3. On a "typical" 34 nm SSD (like Intel X-25M) what's the likelihood of a soft error changing a single bit in my data? Is it something like 1 error per X GB or is it practically zero until NAND degrades after enough use? If I understand correctly, the RAISE error correction graph implies it's a function of time, and the table below the graph shows the probability of uncorrectable errors hitting the end user is significant.
Finally, a suggestion. Data integrity is important to a lot of SSD users (even consumers), and I know you test reliability as much as possible and have shown older SSDs getting bricked or becoming unuseable. Many of us do nightly backups and aren't that scared of total failure. I'm more worried about data being slowly corrupted as a bit changes here or there. Would it be difficult to add to your testing suite the equivalent of memtest for SSDs? It could write and rewrite patterns to the drive constantly for a week or so and see if anything out of the ordinary comes up.
With all the competition I think SSD makers have a incentives to sacrifice "soft error" reliability in favor of price and capacity, and that worries me.
mpx - Friday, December 3, 2010 - link
Variation is suppsed to be very high, with some producers selling "Select" SLC flash models, where 93% of cells can withstand 2 million cycles. It goes into enterprise drives like STEC, while the non-select, but still within spec one one goes to fast Compact Flash or SD cards.Endurance is not the only parameter affected by going analog in MLC. Also suceptibility to electric noise, that may cause misread of a charge level. This is also getting worse with higher densities. Three-bit cells are even worse here.
In enterprise systems when the cells of an SSD or a flash module are near the wear limit the system log of your array informs about this, or even e-mails message about this. Even if it's ignored then there's always redundancy of RAID - SSD will be simply swapped to another, new SSD, that begins wear-process from the start.
More theory on this site:
http://www.storagesearch.com/ssd-slc-mlc-notes.htm...
Mr Perfect - Thursday, December 2, 2010 - link
With the last generation of flash, manufacturers stopped talking about SLC and started pushing Enterprise MLC. Is there some reason SLC is loosing out? I would think that at smaller nodes, SLC would not only become affordable, but might actually start to make sense with it's significantly higher endurance.NandFlashGuy - Thursday, December 2, 2010 - link
Yes, SLC is losing out because of the cost. SLC drives are still available, but they are generally only for extreme usage cases where the usage justifies the cost.Link to Hitachi and Intel's jointly designed enterprise SSD:
http://www.hitachigst.com/solid-state-drives/ultra...
A hard drive in your laptop is written to far less than the capability of MLC flash -- so there's no reason to double the cost for endurance that will never required.
Enterprise MLC is a nice solution for those who demand beyond normal MLC capabilities but not full-fledged SLC pricing.
Longer term, I think you'll continue to see diversification in the market. Perhaps Three-level cell SSDs for consumers who are extremely price sensitive vs higher endurance solutions for those willing to pay more.
melgross - Thursday, December 2, 2010 - link
It seems to me that no matter how sophisticated controllers and memory becomes, the decreasing writability of the cells will overcome all of this. We're dependent on those numbers, and at some point the write cycle max will be reached. I don't quite understand how a controller can get around it. All they can do is to use the cells more efficiently. At some point, there will be no more gains possible, and going to smaller cells will result in decreased lifetime, requiring so much spare capacity on the drive that the point of using smaller process tech will have been negated. Sandforce already uses up to 28% spares on it's better drives. Where will that go at 25nm? What about the less expensive models such as the OWC with 7% I just bought?If they can't get 25nm much better than 3,000 writes, then the next smaller node may be no better than about 1,500. Can they really deal with that? I don't see how.
ABR - Friday, December 3, 2010 - link
Let's assume 100MB/sec is an attainable sustained rate, and generate/write randomized data repeatedly over the whole disk as a single 64GB file and then delete it, and do it again. 64000 / 100 = about 11 minutes for one cycle. It would take a little over 3 weeks to hit a 3,000 limit.Sounds bad. But, my laptop disk has written about 64GB in *2 weeks*: 1800 times slower usage. If it was an SSD it would last a hundred years at that rate.
mpx - Saturday, December 4, 2010 - link
There are some software problems involved. Windows frequently moves a lot of files through main disk drive even if it doesn't need it.. For example when using decompression using copy-and-paste in some decompression programs the files are first decompressed to %TEMP%, and only then moved to final location. This doesn't happen if you use "Extract" from program menu, it's the issue with Windows Explorer.flamethrower - Thursday, December 2, 2010 - link
Hey Anand,Are you able to test endurance? You are not, correct? Because the wear leveling algorithm is too good. You cannot R/W the same 1MB of data 10,000 times because the wear leveling algorithm will spread that wear across the entire drive. You could try randomizing the data as it is written (because the SandForce controller will cheat and say data has been written when it has not actually been if the data is the same or even similar). Even if you create a small partition and write to it over and over, that still will not work. Correct?
Is there a way you can write to the drive in raw mode, testing endurance that way?
Iketh - Thursday, December 2, 2010 - link
ie bypassing the controller?? lol doubt itNandFlashGuy - Thursday, December 2, 2010 - link
One can certainly purchase raw Nand and use a Flash programmer/reader to cycle a small portion of the Nand device and measure Raw Bit Error Rates as a function of cycle.Obviously those in the industry do this all the time.
Here's a link to a often-cited paper comparing Raw Bit Error Rates:
http://ieeexplore.ieee.org/xpls/abs_all.jsp?arnumb...
ProDigit - Sunday, December 5, 2010 - link
Interesting,if higher yields are causing more errors, I might wanna buy Intel 32nm processors NOW before the 25 nm come out with higher error rates, and lower clock speeds!
About the drives, 32nm is already pretty awesome I'd say!
I still own a 72nm ssd somewhere!
compvter - Tuesday, December 14, 2010 - link
Didn't your (or intel's) slide state that 3rd gen ssd would increase life span from 7.5TB - 15TB to 30TB - 60TB. Both have 80gb model as smallest so smallest ssd drives lifespan should go way up (4x).http://www.anandtech.com/show/3965/intels-3rd-gene...
So i wouldn't be that worried if i was planning on buying 25nm ssd.
Kevlion - Friday, December 17, 2010 - link
" today 25nm MLC NAND is good for around 3,000 program/erase cycles per cell.:"I see a lot of statistics on SSD regarding life of an SSD, and I consider myself in the extremely large mass of people still watching the SSD market to find when the value mark hits and I will take the plunge into SSD.
For the vast majority of us, SSD still don't make sense based on $ and increased performance. (Note I fully expect the ANAND readership to disagree on their personal experiences with SSD and how great they are at improving performance of their machine). However plunking down another 25% in a system build and getting 1/4 the size of a standard drive and still having to buy a standard drive as a second in the system just seems like too much at the pace the industry is changing. I mean with my last computer lasting 5 - 7 years (I know there will be those of you who groan at this) I still consider this an average life cycle for most people. That said If I plunk down this now, I need to know that it will last my 5 - 7 years. I don't want to have to re-invest in another SSD in 3 years because of decreased read/write cycles on each new generation.
All this leads to my question: What do these numbers mean to an average person like me that plays games for an hour or two every night, and uses the web for email and news???
To make an informed decision on SSD it's imparitive that we can weigh our own use versus product specs, but these numbers just are too big to make sense of... maybe there is a rule of thumb to use?