SSD Write Endurance 25nm Vs 34nm

[Anvil]

Some sort of POST leds would have been interesting though, it was the same issue as the Samsung for sure.

I'm on the fence right now, changing the test makes it a different test but we can't ignore what happened.
So, maybe a new set of drives and a different test pattern once MWI is exhausted.

[johnw]

I'm on the fence right now, changing the test makes it a different test but we can't ignore what happened.
So, maybe a new set of drives and a different test pattern once MWI is exhausted.

I agree that the currently running drives should not have major changes to the procedure. Maybe a one-week intermission every 100 or 200 TiB would be interesting.

But I am planning on testing a 64GB Samsung 830 or Intel 520 (won't decide which until the 520s are out), and I think I am going to include frequent one-week powerdowns in the test procedure.

[Christopher]

I think it is obvious what happened. Same as happened to the Samsung 470. Once the unpowered data-retention time gets down to a few days, powering off the SSD for several days causes most of the flash to lose its charge. Since the SSD stores critical metadata on the flash, the SSD cannot operate once most of the flash has lost its charge.

And that isn't satisfying. Its more of a lament.

And to think there wasn't a single reallocation before the drive died. It may be a symptom of 25nm NAND... The cells can't hold a charge before they fail. Maybe 34/32nm NAND will die first before the gates get too warn to hold in the voltage.

[zads]

...It just didn't like 9 days without power.

...This data retention issue does however bring an element of frustration to the testing. Is a year unreasonable? How much of an impact does that requirement have on how the MWI is calculated? At what point should you stop writing and for how long should you then wait?....

Maybe something kept the M4 controller from reinitializing on power on? I find it hard to believe that the M4 just quit working after being powered off for a few days....

...Once the unpowered data-retention time gets down to a few days, powering off the SSD for several days causes most of the flash to lose its charge. Since the SSD stores critical metadata on the flash, the SSD cannot operate once most of the flash has lost its charge.

You guys are on the right track here.
The flash manufacturers rate the PE cycle based on flash data retention and correctability given the amount of spare ECC area on each page and how much that can correct. So you can theoretically run 200%, 300% etc of the PE cycle, but you're gonna get more and more probability of errors or failed drives. The bit error probability skyrockets above the 10^(-15) error rate that most consumer drives quote at this point. When a IMFT 25nm drive starts out, it is likely well beyond a 10^(-20) unrecoverable bit error rate.
Throw in things like temperature variations, background radiation, etc, and you have a smattering of other error inducing mechanisms.
The reason its not booting up is likely that the last remaining firmware copy in your flash got corrupted beyond recoverability.

True enterprise SLC and MLC drives actually push more width and more layers of error correction mechanisms into the drive, which ultimately gives them longer data retention and lower unrecoverrable bit error rates than consumer drives, from infancy to EOL.

[B.A.T]

I'm already missing the m4

I know. It's a sad thing to continue without it

[Christopher]

You guys are on the right track here.
The flash manufacturers rate the PE cycle based on flash data retention and correctability given the amount of spare ECC area on each page and how much that can correct. So you can theoretically run 200%, 300% etc of the PE cycle, but you're gonna get more and more probability of errors or failed drives. The bit error probability skyrockets above the 10^(-15) error rate that most consumer drives quote at this point. When a IMFT 25nm drive starts out, it is likely well beyond a 10^(-20) unrecoverable bit error rate.
Throw in things like temperature variations, background radiation, etc, and you have a smattering of other error inducing mechanisms.
The reason its not booting up is likely that the last remaining firmware copy in your flash got corrupted beyond recoverability.

True enterprise SLC and MLC drives actually push more width and more layers of error correction mechanisms into the drive, which ultimately gives them longer data retention and lower unrecoverrable bit error rates than consumer drives, from infancy to EOL.

Does that mean that the M4 would have continued writing had it not been powered off, and might have continued until the NAND started wearing completely out? The M4 didn't even have any reallocation events, but somehow on power-off I guess the cell charge dissipated. I think it's an artifact of 25nm NAND, that 3xnm NAND gets flagged as bad before the drive gets to the point where MLC trapped voltage "leaks out" of the weakened gates. Maybe I'm phrasing it wrong, but the M4 had 13,330PE cycles on it.

I know. It's a sad thing to continue without it

I basically stole my M4 back from the family member I gave it to (giving them another drive in its stead). I'm thinking about OPing it and running it, but I'm not sure how many writes it has already. I would guess it's under 200GB.

I just upgraded it to 0009 (It had 0001FW, and yours had 0002 I think). I'm not sure how much value it would be to the test though. That thing was a beast, but once you get to 500TB you should probably start leaving it off for a couple days. There's no guarantee that a second drive would die the same death, but I'll put it out there.

[Christopher]

Mushkin Chronos Deluxe 60 Update, Day 47
3.3.2FW

05 6 <--Increase by 1, happened today
Retired Block Count

B1 11
Wear Range Delta

F1 465745
Host Writes

E9 359344
NAND Writes

E6 100
Life Curve

E7 10
Life Left

Average 131.02MB/s Avg
Intel RST drivers, Celeron G530 Biostar TH67+

1093 Hours Work
Time 1 month 15 days 13 hours

11 GiB Minimum Free Space

[B.A.T]

I basically stole my M4 back from the family member I gave it to (giving them another drive in its stead). I'm thinking about OPing it and running it, but I'm not sure how many writes it has already. I would guess it's under 200GB.

I just upgraded it to 0009 (It had 0001FW, and yours had 0002 I think). I'm not sure how much value it would be to the test though. That thing was a beast, but once you get to 500TB you should probably start leaving it off for a couple days. There's no guarantee that a second drive would die the same death, but I'll put it out there.

A quick check of value 173 and multiply it with 64 020 803 584 byte and you will have your answer. I think the only impact fw ver have is the speed during the test. Most of the improvement from v 0001/0002/0009 was stability issues. It's a very good ssd and a little strange to continue without it. It's been in this test since 1th july and we need to rethink how we do the testing and to what purpose we are testing. Until now we have just been counting P/E cycles.

[Christopher]

The Mushkin is writing almost 250GB more a day now than it was at the beginning. It was doing about 10.6TB a day, and now it's writing about 10.86TB a day in host writes... not too shabby.

EDIT

I got another reallocation in just the past few minutes.


B.A.T.
The M4 is sitting at all of 4PE cycles, and a whole lot of non-4K aligned accesses.

[Anvil]

...
The reason its not booting up is likely that the last remaining firmware copy in your flash got corrupted beyond recoverability.
...

Are you saying that the firmware is stored in regular NAND?

--

Kingston SSDNow 40GB (X25-V)

459.96TB Host writes
Reallocated sectors : 12

MD5 OK

33.50MiB/s on avg (~84 hours)

--

Corsair Force 3 120GB

01 90/50 (Raw read error rate)
05 2 (Retired Block count)
B1 64 (Wear range delta)
E6 100 (Life curve status)
E7 13 (SSD Life left)
E9 328050 (Raw writes)
F1 436652 (Host writes)

MD5 OK

106.88MiB/s on avg (~84 hours)

power on hours : 1277

B1 has stopped decreasing, it is 2 up from last reading : from 62 to 64.

--

We are not going to change the test, the goal still is to find how long the drive can Endure writing, it is the main goal.

However, as a result of the Samsung and the m4 stopped working there is now a second goal, data retention.
It has been there all along and the m4 just confirms that we need to start looking at the phenomena.

The question is how do we validate the retention without making the test last for years and thus making the test academic and boring.

I am continuing as planned and will perform a short retention test at 500TiB on both drives, testing every 50 or 100TiB depending on speed might be what we need.

Instead of all drives going through rigorous retention testing we should start by selecting a drive (a fast drive e.g. the Samsung 470/830 or the m4) and perform more exact retention testing on that drive.
That way we can get to the core without breaking the main goal of the test.
The findings from this test can be used to specify how to perform a proper retention test.

[Christopher]

Anvil,

If its not too much trouble could you post some benchmarks of the 830 128GB? ASU and CDM or whatever would be excellent. I've only seen the 256 and 512, and since Plextor just released a much faster version of its Marvell/toggle drives I'm a little torn between the two at 128gb.

I'm planning on doing some retention testing, but not until somewhere north of 600TB raw writes since the Mushkin is using 32nm NAND.

[Anvil]

I already posted a thread about the 830

Link to Samsung 830 Benchmarks

[johnw]

Are you saying that the firmware is stored in regular NAND?

Highly unlikely. Since the firmware does not need to be written frequently, it can be stored on its own private flash that is used for nothing else.

In contrast, the metadata for the index of LBAs to flash pages is constantly changing, so it is likely written to regular flash (probably redundant copies, too) so that it can take advantage of wear leveling. If the index were written to private flash, it might wear out before the rest of the SSD's flash.

Although I can think of another strategy for writing the metadata, so I could be wrong about it being written to regular flash. The other strategy would be to devote a chunk of flash many times larger than the metadata. Then do a simple wear-leveling algorithm where each write of the metadata occurs in the next slot in the reserved flash. I'm not sure how much flash would need to be reserved, since it depends on how often the metadata gets written to flash.

[bluestang]

M225->Vertex Turbo 64GB Update: Milestone Reached...

703.05 TiB (773.01 TB) total
1838.56 hrs (Torture), 2641 hrs (Power-On)
13884 Raw Wear
120.03 MB/s avg for the last 22.82 hours (on W7 x64)
MD5 OK
C4-Erase Failure Block Count (Realloc Sectors) at 12.
(1=Bnk 6/Blk 2406; 2=Bnk 3/Blk 3925; 3=Bnk 0/Blk 1766; 4=Bnk 0/Blk 829; 5=Bnk 4/Blk 3191; 6=Bnk 7/Blk 937; 7=Bnk 7/Blk 1980; 8=Bnk 7/Blk 442; 9=Bnk 7/Blk 700; 10=Bnk 2/Blk 1066; 11=Bnk 7/Blck 85; 12=Bnk 4/Blk 3192)

[devsk]

M225->Vertex Turbo 64GB Update: Milestone Reached...

703.05 TiB (773.01 TB) total

21 days to 1PB... If you keep using Windows 7 all the time. Please keep using W7. Its a good OS!...

[RyderOCZ]

I have a feeling that bluestang's drive would suffer the same issue as the M4 if you turn it off.

[bluestang]

I have a feeling that bluestang's drive would suffer the same issue as the M4 if you turn it off.

I'll take that as insider knowledge and just let it keep going

Out of curiosity, if it was off for a few minutes would it kill it?

[Anvil]

Hard to tell, lets not find out the hard way

[B.A.T]

Todays update:
Kingston V+100
202.3251 TiB
799 hours
Avg speed 74.79 MiB/s
AD still 1.
P/E?
MD5 OK.
Reallocated sectors : 00



Intel X25-M G1 80GB
24,7117 TiB Host writes
18762 hours
Reallocated sectors : 00
MWI=65 to 57
MD5 OK
52.56 MiB/s on avg

[B.A.T]

It seems that the MWI might not be so overly conservative after all. 1 year to hold data without a charge after the MWI has expired seems excessive. I wonder how much the MWI would change if the requirement to hold data changed to 3 months as it is with enterprise.

Zads

It is rated at 1 year after hitting MWI-1.

How far was the M4 past the MWI-1?

It hit MWI-1 at 178TiB. It was up to around 600TiB of writes if memory serves.

Way past MWI 1. If it had been stopped at that point it should have held its data as speced.
After that holding data is questionable to length of time.
Same as asking how long it will continue to write data.

You are all making the new specs with the testing.

Do you know where I can find the spec. The M4 spec sheet says nothing.

[Ao1]

JEDEC Standard No. 218
A Solid-State Drive (SSD) Requirements and Endurance Test Method

http://www.jedec.org/standards-docum...e-mmc%E2%84%A2

[RyderOCZ]

I'll take that as insider knowledge and just let it keep going

Out of curiosity, if it was off for a few minutes would it kill it?

Only reason I say that, is because you have the older 50nm Nand, which should have 10K PE rating. You have written 13K+ to the whole of the drive, thus you have exceeded the number of PE cycles that would be within the "guaranteed for 1 year retention" spec.

[Christopher]

I have to believe that the VT will last much longer while retaining data than the M4. The M4 had over 4x its NANDs PE rating, while the the Turbo is only 30% past rated PE cycles. I'd be shocked if it didn't make it to 1 PB.

Incidentally, that is why I picked up a VTurbo 120 a few weeks ago.

My older Agility 60 died (at the most inopportune time) today as well, making it my first SSD failure. I was a little surprised as its not like it had a lot of writes on it, but I'm thinking it has more to do with the 1.7fw... I think I'm sticking with 1.6 on the other Indilinxes. I just needed it to not die until Thanksgiving. It will be a few days before I get back to the homestead to check it out, so I'm unable to determine whether it's a d-flash candidate or j.f.d.

[johnw]

I just needed it to not die until Thanksgiving.

Is "it" a turkey?

[Christopher]

Is "it" a turkey?

All I know is, something fowl is afoot.

ouch