Gene said:
That answer is dated 2004. Maybe we can contact the poster
(crabcakes-ga) and find out how his memory sticks are doing. We should
be able to get meaningful statistics in a seven year test
I tend to be suspicious of "accelerated tests". I wonder how accurate
the extrapolation from test conditions to real life can be. That's why
on some level my above remark isn't quite a joke.
OTOH, I wonder if today's flash memory cards and flash drives might be
more resilient. Seven years of R&D...
Ten years is a more or less reasonable answer. There is a risk of "bit rot"
after enough time has passed.
KFW4G16Q2M - SLC - "Endurance : 100K Program/Erase Cycles Data Retention : 10 Years"
http://www.samsung.com/global/syste...Gbit/KFW4G16Q2M/ds_kfxxg16q2m_66mhz_rev12.pdf
For the following one, I got a copy from a datasheet site, as Samsung wants
an NDA for it. The download mechanism means it's difficult to give a direct
link to the datasheet (K9G8G08U0M__datasheet_co_kr.pdf)
K9G8G08U0M - MLC - Reliable CMOS Floating-Gate Technology
Endurance : 5K Program/Erase Cycles(with 4bit/512byte ECC)
Data Retention : 10 Years
So this is not in the same class as say, archival grade optical media with
a claimed life of 100 years. No attempt is made to over-promise this stuff.
Bit rot is occasionally seen, but the observation is on NOR flash, rather
than NAND. Since NAND is typically covered by ECC, you don't get to see
how many bits are really in error. Current NAND chips will ship with some
bits already bad inside, and the ECC hides it.
NAND is headed in the wrong direction, at least in the case of MLC. Whether
the manufacturers will insist on SLC doing the same thing, is another question.
In order to sell "enterprise grade solutions", SLC has to be demonstrably better.
Some MLC flash now, only has a 3K write cycle life - like hard drives, MLC
is being pushed for all it's worth, in the "density" direction, at the
expense of virtually any other parameter that get in the way. This means
cranking up the ECC code, to cover more bit errors, to make up the
difference.
But write cycle life wasn't the original question, and as far as data retention,
there seems to be no change in the boilerplate value. Naturally, they don't
actually wait ten years, to see if any bits change. This is determined by
"accelerated life testing", with the assumption that the modeling is understood
well enough, that the acceleration method is representative of real devices.
You can have flash memory, erased by ionizing radiation sources. So there are
mechanisms other than those planned in the datasheet, that could have an impact.
http://cmoset.com/uploads/Marta_Bagatin_2010.pdf
"Effects of ionizing radiation are becoming more and more severe on
Floating Gate cells. Only few hundred electrons separate adjacent program
levels. In advanced devices (< 65 nm) atmospheric neutrons can induce errors
(without ECC)"
Up to a certain level, ECC can hide bad bits or bits corrupted during storage
lifetime. But then, if you can't access the uncorrected state of the device,
then you don't know how bad it's getting. With a CD or DVD, you can do an
error scan, and "see" the rot happening over time. I don't know if that is
possible or practical with Flash memory devices or not.
Paul