MMM
Page 1 of 2 12 LastLast
Results 1 to 25 of 45

Thread: The life expectancy of SSD's

  1. #1
    Xtreme Mentor
    Join Date
    Feb 2009
    Posts
    2,597

    The life expectancy of SSD's

    This post is to try and understand the life expectancy of SSD's.

    The most informative information comes from Intel. The following references have been used to obtain other assumptions:

    http://techreport.com/articles.x/15433
    http://www.anandtech.com/cpuchipsets...spx?i=3403&p=4

    Intel are quoted in these articles as stating that OEM's required them to prove that their SSD could provide 20GB writes per day for 5 years.

    The variable factors on an SSD architecture are write amplification and wear levelling. (Cell construction is not considered and input on this would be welcome.)

    Intel are quoted as stating that they use the following formula to calculate this.
    Cycles = (Host writes) * (Write amplification factor) * (Wear levelling factor) / (Drive capacity)

    Intel are quoted as claiming that other SSD manufactures use this formula:
    Cycles = (Host writes) / (Drive capacity)

    Based on a 32GB drive the difference between formula's give a variation of approximately 40%.

    Using Intel's figures the life expectancy can be calculated as follows:

    Intel formula based on 32 GB drive @ 20GB per day =
    0.715 = (20) x (1.1) x (1.04)/ 32
    0.715 cycles x 1,825 (5 years) = 1304.87
    (Total life cycles) 10,000/ 0.715 cycles = 13,986 days or approx 38 years.
    Drive will last approximately 38 years writing 20 GB per day

    Intel formula based on 32 GB drive @ 100 GB per day =
    3.575 = (100) x (1.1) x (1.04)/ 32
    (Total life cycles) 10,000/ 3.575 cycles = 2,797 days or approx 7 years, 6 months
    Drive will last approximately 7 1/2 years writing 100 GB per day

    Intel formula based on 80 GB drive @ 20 GB per day =
    0.286 = (20) x (1.1) x (1.04)/ 80
    0.286 cycles x 1,825 (5 years) = 521
    (Total life cycles) 10,000/ 0.286 cycles = 34,965 days or approx 95 years, 10 months
    Drive will last approximately 95 years writing 20 GB per day

    Intel formula based on 80 GB drive @ 100 GB per day =
    1.43 = (100) x (1.1) x (1.04)/ 80
    1.43 cycles x 1,825 (5 years) = 2,609.75
    (Total life cycles) 10,000/ 1.43 cycles = 6,993 days or approx 19 years
    Drive will last approximately 19 years writing 100 GB per day

    Using Intel's formula and Intel's assumptions on other brands of "mediocre" quality" SSD's with a write-amplification factor of 10 and a wear-levelling efficiency factor of 5 (ref Tech Report) you get the following results:

    32GB drive @ 20GB per day. Assume write amplification of 10 & wear levelling at 5
    31.25 = (20) x (10) x (5)/ 32
    31.25 cycles x 1,825 (5 years) = 57,031
    (Total life cycles) 10,000/ 31.25 cycles = 320 days or approx 1 year
    Drive will last approximately 320 days writing 20 GB per day

    32GB drive @ 100GB per day. Assume write amplification of 10 & wear levelling at 5
    156.25 = (100) x (10) x (5)/ 32
    156.25 cycles x 1,825 (5 years) = 285,156
    (Total life cycles) 10,000/ 155.25 cycles = 64 days
    Drive will last approximately 64 days writing 100 GB per day

    As can been seen the estimated life time is heavily dependent on the following factors if you assume 20GB writes per day:

    • Size of the drive
    • Write amplification factor
    • Wear levelling

    EDIT: Wear levelling:

    Static wear leveling
    This technique wear-levels over both dynamic and static data areas, so the entire capacity of the SSD can be used for wear leveling.
    Reassigning static data is a more complex task than dynamic data because it requires multiple operations to safely move static data around. As a result, this implementation may impact the overall write performance.

    Dynamic wear leveling
    This technique only wear-levels over dynamic data areas. For example, an 30GB SSD with 20GB of dats only leaves the remaining 10GB capacity for wear leveling. Dynamic wear leveling has a shorter life expectancy compared to static wear leveling because only dynamic data areas can be used to rotate data, but write times improve.
    Last edited by Ao1; 04-30-2009 at 01:20 AM.

  2. #2
    Xtreme Member
    Join Date
    Jun 2005
    Location
    Germany
    Posts
    427
    Lifetimes of under a year would be pretty bad for us, but worse for the company producing them. Imagine the high number of returns since you have at least one but mostly two years of warranty. So i can't imagine that they will last under the time of warranty, maybe if you're running these benchmarls all day and trying to kill the drive

    Xeon W3520 D0
    Gigabyte GA-EX58-UD5
    Powercolor HD 6950
    Areca 1680ix-16
    3x Vertex 60GB @ Raid 0
    5x 5900.12 1.5TB @ Raid 5
    X-Fi Xtreme Music
    Coolermaster UCP 700W


    EK Res 250 -> 2x DDC-1T w. XSPC Dual Top -> EK HD6970 -> EK Supreme HF Full Nickel -> MoRa3 (9x140)

  3. #3
    Xtreme Mentor
    Join Date
    Feb 2009
    Posts
    2,597
    Quote Originally Posted by ic3m4n2005 View Post
    Lifetimes of under a year would be pretty bad for us, but worse for the company producing them. Imagine the high number of returns since you have at least one but mostly two years of warranty. So i can't imagine that they will last under the time of warranty, maybe if you're running these benchmarls all day and trying to kill the drive
    I think the figures have to be taken with a degree of caution, but none the less they put something on the table for discussion.

    JS Monitor can estimate the erase count and in other forums it is showing close to the total erase count so no doubt things will become a lot clearer if theory can be matched with practice.

  4. #4
    SLC
    Join Date
    Oct 2004
    Location
    Ottawa, Canada
    Posts
    2,795
    Anyone up for buying a small jmicron drive, filling it with IOMeter test file and writing to it 24/7 at max sequential write speed until it stops working?

    It can write at about 8200GB/24 hours. Writing sequentially would make the write amplification a lot smaller though so it would take longer to die, but it should still be within a few days according to those formulas.

  5. #5
    Xtreme Addict
    Join Date
    Mar 2008
    Posts
    1,163
    1. Home users don't write 100 or even 20 GB every day. I doubt that more than 0.01% write more than 1 GB on average.
    2. 10K writes per cell means that 98% cells with last at least that long. Some longer, even much longer. In case of Texas Memory Systems SLC up to 5 times declared life and overall more than what's stated.
    3. this and this.
    4. I wouldn't trust Intel marketing claims referring to competing products.

  6. #6
    SLC
    Join Date
    Oct 2004
    Location
    Ottawa, Canada
    Posts
    2,795
    I am ALMOST curious enough to spend 100 bucks on a drive just to kill it.

  7. #7
    Registered User
    Join Date
    Oct 2006
    Location
    Kirghudu, Cowjackingstan
    Posts
    462
    My view on this is following: if it dies within warranty - I don't care. If it dies shortly after... Well then, it served me well and it cost me not more than Purchase price/24 per month.

    Quote Originally Posted by One_Hertz View Post
    I am ALMOST curious enough to spend 100 bucks on a drive just to kill it.
    Do it. $75 for 16Gb.
    Last edited by F@32; 03-25-2009 at 08:23 AM.

    Sony KDL40 // ASRock P67 Extreme4 1.40 // Core i5 2500K //
    G.Skill Ripjaws 1600 4x2Gb // HD6950 2GB // Intel Gigabit CT PCIe //
    M-Audio Delta 2496 // Crucial-M4 128Gb // Hitachi 2TB // TRUE-120 //
    Antec Quattro 850W // Antec 1200 // Win7 64 bit

  8. #8
    Xtreme Mentor
    Join Date
    Feb 2009
    Posts
    2,597
    Hmm Intels limited warranty states:

    EXTENT OF LIMITED WARRANTY

    Intel does not warrant that the Product will be free from design defects or errors known as “errata.” Current characterized errata are available upon request.



    Note sure what errata covers, I will ask them for clarity.

  9. #9
    SLC
    Join Date
    Oct 2004
    Location
    Ottawa, Canada
    Posts
    2,795
    Quote Originally Posted by m^2 View Post
    1. Home users don't write 100 or even 20 GB every day. I doubt that more than 0.01% write more than 1 GB on average.
    2. 10K writes per cell means that 98% cells with last at least that long. Some longer, even much longer. In case of Texas Memory Systems SLC up to 5 times declared life and overall more than what's stated.
    3. this and this.
    4. I wouldn't trust Intel marketing claims referring to competing products.
    1. An untweaked OS itself if left idling will write a few GB a day. Somebody somewhere did some calcs and it was like 4GB/day with very light firefox use. I wish I remembered where I saw that. This may be FUD, my memory is fuzzy.
    2. This is actually questionable... There are some forensics experts that say that those cells don't last anywhere near their rated 10k/100k values, but actually much less. It makes sense that manufacturers would lie about this, and the forensics guys have no reasons to lie, so I am keeping my doubts on those numbers for the time being. I will post a link of one expert mentioning this in a presentation this evening.
    3. I've received some more dead sammys since I made those posts. All had burnt out controllers and the NAND chips were actually just fine and perfectly readable with specialized and very expensive NAND chip reader hardware.
    4. Yep. Probably over exaggerated.

  10. #10
    Xtreme Mentor
    Join Date
    Feb 2009
    Posts
    2,597
    Quote Originally Posted by m^2 View Post
    1. Home users don't write 100 or even 20 GB every day. I doubt that more than 0.01% write more than 1 GB on average.
    2. 10K writes per cell means that 98% cells with last at least that long. Some longer, even much longer. In case of Texas Memory Systems SLC up to 5 times declared life and overall more than what's stated.
    3. this and this.
    4. I wouldn't trust Intel marketing claims referring to competing products.
    20GB is what the OEM's required. No doubt over the top, but I guess an OS like Vista is writing all the time in the background. What that adds up to I have no idea.

    Right on regarding the marking claims, but it was all I could find to go on. No one else seems to publish this type of data.

  11. #11
    Xtreme Member
    Join Date
    Jan 2009
    Posts
    165
    It is hard to believe that if you can kill a mcl drive in a couple of days nobody has tried it yet

    Lets say you have a 32gb drive and it's 99% full, you could kill the remaining free space in a matter of hours
    Last edited by annihilus; 03-25-2009 at 08:59 AM.

  12. #12
    Xtreme Addict
    Join Date
    Mar 2008
    Posts
    1,163
    Quote Originally Posted by One_Hertz View Post
    1. An untweaked OS itself if left idling will write a few GB a day. Somebody somewhere did some calcs and it was like 4GB/day with very light firefox use. I wish I remembered where I saw that. This may be FUD, my memory is fuzzy.
    Silicon Systems and MTRON have a tools that measure wear of their drives. Maybe others too. This is probably the most reliable way because it measures real, not estimated write amplification factor.

    Quote Originally Posted by One_Hertz View Post
    2. This is actually questionable... There are some forensics experts that say that those cells don't last anywhere near their rated 10k/100k values, but actually much less. It makes sense that manufacturers would lie about this, and the forensics guys have no reasons to lie, so I am keeping my doubts on those numbers for the time being. I will post a link of one expert mentioning this in a presentation this evening.
    I'll gladly take a look.

    Quote Originally Posted by annihilus View Post
    It is hard to believe that if you can kill a mcl drive in a couple of days nobody has tried it yet

    Lets say you have a 32gb drive and it's 99% full, you could kill the remaining free space in a matter of hours
    Not with smart wear levelling. Some move pages that are read only to get more even wear.

  13. #13
    Xtreme Mentor
    Join Date
    Feb 2009
    Posts
    2,597
    Quote Originally Posted by m^2 View Post
    Silicon Systems and MTRON have a tools that measure wear of their drives. Maybe others too. This is probably the most reliable way because it measures real, not estimated write amplification factor.
    That would really help on the speculation. I keep looking at those Intel figures to see if I have made a mistake. The durations seem way too long.

    With regards to monitoring tools I think you have to have the JS Monitor running all the time so that it can calculate the erase count, which does not help if you have not used it consistently.

    Do you have any info on the Silicon Systems/ Mitron tools?

  14. #14
    Xtreme Addict
    Join Date
    Mar 2008
    Posts
    1,163
    Every drive remembers erase count of each block, it's needed for wear levelling. These tools simply read it from the drives. I don't remember the names of these programs, but that shouldn't be hard to find.

  15. #15
    Xtreme Member
    Join Date
    Jun 2005
    Location
    United Kingdom of Great Britain and Northern Ireland
    Posts
    464
    There is such a tool for JMicron drives, but nobody seems to be aware of any for Indilinx. As for Intel I haven't heard anyone mention one - but I'm not as fluent with their drives. If Intel is suggesting their write amplification is 1.1 compared to everybody else at 10, then I simply don't believe them. The minimum possible is 1. Suggesting that Intel have managed to make theirs 100 times better than the competition sounds a little bit optimistic on Intel's part.

    I'm quite sure it's possible for people to measure the amount of writes to a drive over the course of a week, and from there we can better tell how things are. 20GB sounds like an awful lot, however I wouldn't rule it out. When I've watched what my OS is doing in the background it's amazing the amount of registry entries that go on and the amount of seemingly pointless file accesses etc that go on. That's from the Operating System itself, who knows what it could be like from some software like Quicktime, Realplayer, Acrobat etc which don't give a crap about performance as long as they advertise to you constantly. What about if someone has a piece of spyware installed and that causes large amounts of writes over a period of time. Are we looking at people potentially cutting years off the lifespan of the drives without knowing it? I guess it is possible. What about people with low memory - hence much more swap file activity. Is someone with 2GB of RAM going to have their SSD last 1/10th of the time of someone who has 4GB? Again that's possible.

  16. #16
    SLC
    Join Date
    Oct 2004
    Location
    Ottawa, Canada
    Posts
    2,795
    Quote Originally Posted by Halk View Post
    There is such a tool for JMicron drives, but nobody seems to be aware of any for Indilinx. As for Intel I haven't heard anyone mention one - but I'm not as fluent with their drives. If Intel is suggesting their write amplification is 1.1 compared to everybody else at 10, then I simply don't believe them.
    They didn't say that all others have 10, just some bad ones. It is a marketing trick. They are certainly not 100 times better but the fact that all other manufacturers completely refuse to even state their write amplification and wear leveling factors says a lot too.

    I wouldn't be too surprised if jmicron drives had 10...

  17. #17
    Xtreme Mentor
    Join Date
    Feb 2009
    Posts
    2,597
    Found this on Mtron SLC specs:

    http://www.mtron.net/upload_Data/Spe...IF1_rev0.1.pdf

    3.4 Reliability Characteristics
    3.4.1 Wear-leveling
    Mtron SSD supports both static and dynamic wear-leveling. These two algorithms guarantee the use of all flash memory at the same level of write/erase cycles to improve lifetime limitation of NAND based storage.
    3.4.2 Endurance
    Table 5 Endurance
    Write At 32GB capacity: >140 years @ 50GB write/day
    Read Unlimited
    Note:
    1. The above calculation is based on the guaranteed 100,000 program and erase cycles of flash memory from vendors and the assumption that the write is performed in sequential manner.

  18. #18
    Xtreme Mentor
    Join Date
    Feb 2009
    Posts
    2,597
    Hagiwara offer a SSD Service-Life Assessment Service, but you have to buy one of their drives to get your hands on it. Still if they can do it I guess Intel or anyone else for that matter should be able to do it.

    http://www.hsc-us.com/Embedded/ServLife.htm

  19. #19
    Xtreme Mentor
    Join Date
    Feb 2009
    Posts
    2,597
    Interesting. This link takes you to a PPP from the University of Tokyo. Seems there are two types of wear levelling, dynamic & static. Dynamic only introduces wear levelling on empty dynamic data, static introduces wear levelling over all data, including static data. Static data being OS data. Dynamic being user data.

    http://www.lsi.t.u-tokyo.ac.jp/Takeu...NRIA2008_7.pdf

  20. #20
    Xtreme Mentor
    Join Date
    Feb 2009
    Posts
    2,597
    Quote Originally Posted by F@32 View Post
    My view on this is following: if it dies within warranty - I don't care. If it dies shortly after... Well then, it served me well and it cost me not more than Purchase price/24 per month.
    Just trying to find a way to make sure your coffee beans are fresh

  21. #21
    Xtreme Mentor
    Join Date
    Feb 2009
    Posts
    2,597
    This is a link to a joint white paper by MS & the Univ of Wisconsin called "Design Tradeoffs for SSD Performance" and uses a Samsung SLC for analysis.

    http://www.usenix.org/events/usenix0...tml/index.html

    I think it is saying that dynamic & static wear leveling decreases performance but in different terminology....I need to read it again with fresh eyes to be be sure.

    The cost of migrating cold pages across blocks imposes a performance cost that is workload-dependent. Our simulation of wear-leveling for IOzone involved 7902 migrations per package which added a 4.7% overhead to the average I/O operation latency.

  22. #22
    SLC
    Join Date
    Oct 2004
    Location
    Ottawa, Canada
    Posts
    2,795
    Quote Originally Posted by m^2 View Post
    I'll gladly take a look.
    http://www.youtube.com/watch?v=LY36S...eature=related

    Interesting bits of info there 0:30-0:55ish and 3:25-4:20. The whole 1h presentation is on youtube and is good if you are into this stuff.

  23. #23
    Xtreme Addict
    Join Date
    Mar 2008
    Posts
    1,163
    Quote Originally Posted by One_Hertz View Post
    They didn't say that all others have 10, just some bad ones. It is a marketing trick. They are certainly not 100 times better but the fact that all other manufacturers completely refuse to even state their write amplification and wear leveling factors says a lot too.

    I wouldn't be too surprised if jmicron drives had 10...
    No wonder that others don't give any numbers, they depend heavily on usage patterns. I'm pretty sure that with knowing exact algorithm used by X-25M one could force write amplification factor of 500 or more. By random-like 512B writes that are emptying pages in blocks evenly.
    That's why there are tools to estimate lifetime, it's the only reliable way.

    Quote Originally Posted by One_Hertz View Post
    http://www.youtube.com/watch?v=LY36S...eature=related

    Interesting bits of info there 0:30-0:55ish and 3:25-4:20. The whole 1h presentation is on youtube and is good if you are into this stuff.
    Thank you, I'll watch it tomorrow.

  24. #24
    SLC
    Join Date
    Oct 2004
    Location
    Ottawa, Canada
    Posts
    2,795
    Quote Originally Posted by m^2 View Post
    No wonder that others don't give any numbers, they depend heavily on usage patterns. I'm pretty sure that with knowing exact algorithm used by X-25M one could force write amplification factor of 500 or more. By random-like 512B writes that are emptying pages in blocks evenly.
    That's why there are tools to estimate lifetime, it's the only reliable way.
    Write amplification sure, but wear leveling factor not exactly... It is just how evenly the writes are distributed amongst all cells. It won't (shouldn't if the algorithm is decent!) vary too much. Certainly not from 4% to 400%.

  25. #25
    Xtreme Mentor
    Join Date
    Feb 2009
    Posts
    2,597
    This is an interesting read.

    Extracted from here: http://www.storagesearch.com/ssd-testart.html

    Sometimes the headline numbers get better, sometimes they get worse. There are many good reasons for this, because unlike memory devices, SSDs are complex systems in which software and controller hardware can all play a part in shaping the characteristics of a device. Tweaks in the controller algorithms can have any of the following major effects:-

    • Make the flash SSD go faster.
    • Deal more effectively with endurance
    • Make the data in the flash SSD more reliable / less corruptable.

Page 1 of 2 12 LastLast

Bookmarks

Bookmarks

Posting Permissions

  • You may not post new threads
  • You may not post replies
  • You may not post attachments
  • You may not edit your posts
  •