Break Points
Forgettable Flash
Jack Ganssle
6/15/2009 1:30 PM EDT
We all know about wearing flash out through excessive writes. Plenty of wear-leveling approaches minimize the problem. But will they remember forever?
No. Manufacturers do specify data retention figures for flash. Typically they guarantee ten to twenty years, though the data suggests that flash is a lot better than that.
(It can be hard to get numbers. Micron's web site requires an NDA to get a datasheet! Toshiba's site is nearly unusable, and directs datasheet inquiries to local sales reps. The web is all about making your product information widely available, not gating access through the slow and frustrating process of involving some sales guy.)
TI has an excellent app note about data retention in the MSP430. In it they show that, for users, the most important data retention factor is temperature. There's an inverse exponential relationship between the two.
The good news is that typically these devices will still preserve their programming long after we have all passed on. But at extreme temperatures, be wary.
However, I never use numbers specified as "typical." Worst-case, or guaranteed, figures are the only ones designers can rely on.
TI specs their MSP430F11x1A part with a guaranteed retention time of 100 years. That's long enough for pretty much any application I can imagine. But Freescale specs ten to twenty years. Other vendors have different figures.
Ten years is a long time. I can't imagine anyone using a 2009-era iPod in 2019 (besides, their batteries will be long gone by then). The latest smart phone will look like Ma Bell's black rotary dial unit a decade hence.
But a lot of embedded apps run quietly in the background controlling all sorts of industrial processes for many years on end. A decade just isn't long enough.
We had this problem before. EPROMs were initially guaranteed for ten years. I worked on an 8008-based system in the 70s that went into factories. By the 80s some of these were still in use, and the EPROMs were dropping bits. No one else remembered how to use the development system: an Intellec 8 with ASR-33 paper tape. Though I no longer worked at that company they tracked me down, and I made a nice consulting fee reprogramming the devices.
What about your systems? Is a decade long enough? If not, what actions do you take?
( The Embedded.com Poll Question this week is "What's your product lifetime?" To vote, go to the poll location on the Embedded.com Home Page.
Jack G. Ganssle is a lecturer and consultant on embedded development issues. He conducts seminars on embedded systems and helps companies with their embedded challenges. Contact him at jack@ganssle.com. His website is www.ganssle.com.
|
|
|
|
|
B Kockoth
6/16/2009 4:23 AM EDT
Despite scrapping incentives left and right, automotive flash has to hold 20 years, and after that (think luxury vehicles) it gets really scary, since PC platforms change completely during that time (parallel port, anyone?), so the platform to run the tool is often more of a problem than the historic knowledge how to run them.
When automotive ECUs are dropping bits, effects range from hardly noticable to catastrophic, hence selftests have to run often to keep the system healthy.
Sign in to Reply
BobDJr
6/16/2009 9:31 AM EDT
We were having spurious bit-flip problems in flash a few years ago in modules that were only a year or two old. Even though they were in "unused" memory, the self-test had to check all locations due to the nature of the application. We tried to get to the root of the problem but had to settle for replacing the flash.
The platforms issue consumed a lot of my time at my last job, since much of the code we supported was written in the 80s on the VAX, and before that on things like the Tektronix 8560 MDS.
Fortunately, I'm now on the bleeding edge of nanotech so upgrades take care of these kinds of issues.
Sign in to Reply
kicker24
6/16/2009 12:08 PM EDT
If it could be made cheaper by reducing the retention time to only a year - one must do it. It will save all the industry, many jobs etc.
I guess, that is rather difficult if not impossible. To check the data sheet from a supplier is not necessary if one know how it's measured. Nobody wait 10 years to make a stamp. One uses higher temperature and voltage and extrapolates data to the standard conditions. So one would better speak about the probability to fail, not the time span. This probability increases over time who knows how? Because of that 10 100 or 200 years are always the same thing. Only if you go down to a year, then it's getting interesting...
Sign in to Reply
krwada
6/16/2009 1:09 PM EDT
Amen to that brother!
ok ... about NVRAM, or FLASH. I know of quite a few products out there that appear to be well beyong their life-cycle. The various vehicle sensors that are used in the toll roads back East, and bridges out West. I am talking about things such as FastPass, EZPass, Fastrack etc... How do I know? I am the dude that designed these systems. I can tell you that they all have non-volatile memory in them. NVRAM is a must-have for systems such as these.
I have worried a little bit about the lifecycle retention .... nah! I really have not worried at all about them. I figure that most of the time, these systems are working 24-7, and if anything goes wrong, they usually have spares sitting in some shed that is close by.
... or something like that!
Sign in to Reply
Literate
6/16/2009 4:28 PM EDT
Arrhenius equation for evaluation of the stored charge degradation is typically used by the process people in the cell and arrays developments for extrapolation:
>>http://en.wikipedia.org/wiki/Arrhenius_equation>>
When you get your chip you never know, unless you ask, which technology is behind, what is the stack made of, tunnel dielectric thickness, tunneling mechanisms, read/write voltages and times, amount of charge stored, etc... That is why you never know the real cycling numbers and data retention times.
Another important question - is that NOR or NAND? Since NAND is more vulnerable for any disturbs, especially MLC types and has usually shorter data retention times and durability it is mostly used for mass storage (photos, videos and so forth) whereas NOR is mostly for code/firmware storage.
What can be done about it - if you seriously need to know all this, you have to ask your supplier for the reliable information, if necessary after an NDA, or buy a couple of lots and evaluate it yourself. Here we are coming back to stoves, burn-in tests and the Arrhenus equation. For this you need a specified lab and people.
Sign in to Reply
4ever
6/17/2009 9:06 AM EDT
Dear Mr. Ganssle,
The reliablity of embedded flash is a very important (and luckily, interesting topic). I have been dedicating myself to the reliability of embedded NVM technologies for last 15 years (and currently still working on it).
As "Kicker24" and "Literate" already NVM reliability is not reprensented by a single number (of write/erase cycles, or number of years). In addition, the Arrhenius equation does not cover all the failure modes related to data retention (some mechanisms have 0 temperature acceleration or even negative).
I expect to meet more people like you at the upcoming IEEE international integrated reliability workshop (WWW.IIRW.ORG, in October at Lake Tahoe)to have some up to date technical discussions.
Very best regards,
Sign in to Reply
Andrew B
6/17/2009 9:40 PM EDT
An important correction
- TI specifies 100 years of data retention at +25C, which is meaningless for industrial equipment.
- Freescale specifies 10 years of data retention at FULL industrial temperature range.
So far I have not seen other manufacturer specifying data retention at full industrial temperature range.
Sign in to Reply
vapats
6/18/2009 7:57 PM EDT
JeffL makes a good point, but I must disagree with "where the size of the memory used is too small to make it practical to use error correction techniques"; ECC can be applied to any size/organization of memory, and should be.
In fact, it should be applied to CPU designs as well -- when do *you* want your steer-by-wire system to suddenly flip you into the ditch, because of a cosmic-ray hit?
Sign in to Reply