SSD Write Endurance 25nm Vs 34nm

[Anvil]

Kingston SSDNow 40GB (X25-V)

532.07TB Host writes
Reallocated sectors : 05 14
Available Reserved Space : E8 99

MD5 OK

33.16MiB/s on avg (~85 hours)

--

Corsair Force 3 120GB

01 94/50 (Raw read error rate)
05 2 (Retired Block count)
B1 47 (Wear range delta)
E6 100 (Life curve status)
E7 10 (SSD Life left)
E9 489474 (Raw writes) ->478TiB
F1 651488 (Host writes) ->636TiB

MD5 OK

106.38MiB/s on avg (~85 hours)

power on hours : 1866

--

@B.A.T and Christopher
Make sure to make/save screenshots of SMART on the new drives, if you haven't done so already

[Christopher]

I have pre testing SMART pics and made one corresponding to the first update. I can include them if you want when I post the updates, I just don't know how much bandwidth I have from my box.com account which I use for posting pictures since the forum is broken still.

[B.A.T]

Kingston SSDNow 40GB (X25-V)
@B.A.T and Christopher
Make sure to make/save screenshots of SMART on the new drives, if you haven't done so already

I save screenshots every time for all my ssd in the test. That folder is starting to get very large

[Christopher]

When your really fancy Toggle NAND drive only lasts for 50 days, it helps to keep the number of screenshots down.

I hope the Samsung doesn't last forever.... I only have 50GB of space on my file sharing serivce.........

I bought one of those HP TouchPads when they were being sold for for $99 instead of $500 and you got a free box.net account upgraded to 50GB of storage with it. You get statistical stuff and the ability to direct link downloads and larger files/speeds/etc. You can get the 50GB free now with an Apple device but the extra stuff is, well, extra. I really wish they'd fix the attachments for this thread alone though. It made everything a lot easier.

[Anvil]

You dont have to post/upload the SMART screenshots, just keep a private/local copy until the spredsheet is ready.

[Anvil]

Kingston SSDNow 40GB (X25-V)

533.36TB Host writes
Reallocated sectors : 05 14
Available Reserved Space : E8 99

MD5 OK

36.48MiB/s on avg (~10 hours) ssdtoolbox cleaning + restarted last night as avg was getting low

--

Corsair Force 3 120GB

01 88/50 (Raw read error rate)
05 2 (Retired Block count)
B1 47 (Wear range delta)
E6 100 (Life curve status)
E7 10 (SSD Life left)
E9 492378 (Raw writes) ->481TiB
F1 655353 (Host writes) ->640TiB

MD5 OK

106.40MiB/s on avg (~95 hours)

power on hours : 1876

[Christopher]

I'm pretty sure the Samsung 830 uses 27nm Toggle NAND, 32Gbit dies. So the 64GB has eight packages, dual 32Gbit dies.

Toshiba has 24nm Toggle NAND already shipping in production drives, but they're probably not really the same stuff.

[minpayne]

I'm not quite following up with this thread. Is there any impact by using non-compressible data to fill the SSD? (i.e. would certain controllers "cheat" with 0-fill and increase endurance?) What standard (settings) is used to join this thread?

[Christopher]

I'm not quite following up with this thread. Is there any impact by using non-compressible data to fill the SSD? (i.e. would certain controllers "cheat" with 0-fill and increase endurance?) What standard (settings) is used to join this thread?

The compressibility of data only matters for SandForce drives. They're the only controllers that can compress data/deduplicate data (there may be an enterprise controller made by some other company that does so as well, but I'm not sure. I'm mostly concerned with consumer drives.)

For the SandForce drives we've been using the 46 percent compressibility, but it's pretty much irrelevant for other drives/controllers. In terms of endurance gained by 0-fill for SandForce drives, I think it's six-of-one, half-dozen-the-other. Since the work never ends, you just end up with faster writes of more compressed stuff...

For example, the Mushkin Chronos Deluxe 60GB I had tested would write 90MBs - 100MBs with random data, and around 290MBs with 0-fill (with the endurance testing in ASU). If it's writing 290MBs with 0 fill, the actual writes to NAND are 100MBs. SandForce drives have a SMART value for host writes and a value for raw writes. If you divide the raw writes by host writes you can see the average compression ratio for all the data written to the drive. So a SandForce drive can't cheat in that sense -- you always know how much data is actually written to the NAND after the data has been compressed. At the 46 percent compression ratio in ASU I ended up with about 77 percent of host writes actually being written to NAND. It wouldn't be realistic to only use incompressible data as much of the data being written to the drive is compressible to one extent or another (but nowhere near 0 fill compressible). As a result, you end up with much lower than 1 write amplification, but modern drives are pretty good when it comes to WA.

I tested the Mushkin Chronos Deluxe SF2281 and it died pretty quick. I'm now testing a Samsung 830 64GB, but I just started it about 37 hours ago. If you're serious about testing a drive, I recommend testing a smaller-capacity drive. The two Intel 40GB drives started 7 months ago, and they're still going and going and going. My SF2281 made it 50 days, but that was kinda anomalous. There's no way to say how long a drive will last in advance, but I'd plan on the test lasting several months. I built a low power Sandy Bridge system that I can tuck out of the way in my apartment. I use a remote desktop application for my tablet to use as a monitor. The testing system will need to run 24/7, so having a quiet system is mandatory for me. But ASU doesn't require a lot of resources to endurance test, so if you're a folder you could probably use a folding rig. You could also use a laptop and just close the lid (disable close lid = sleep, never sleep) and stick it under your couch. I tried that for a while. Windows 7 isn't required, but TRIM helps a lot with average speed.

[Christopher]

Samsung 830 64GB Update, Day 2
FW:CXM01B1Q

GiB written:
18185.44

Avg MB/s
107.92

Per-day average:
9092.92



The Samsung starts the loop quite fast, but seems continually drop after the first several seconds, so that at the end of the loop it's only managing 120MBs, and if I let it, it would plunge down into the 100MBs range. It can hit max sequential speeds of 520 read and 160 write which makes it as fast as a 128GB OCZ Octane. I'm not certain that there is a compelling reason to get the 470 instead of the 830. It's just an impressive drive.


Toggle NAND must be made of pixie dust and unicorn glitter.

[bluestang]

@Christopher...your running 46% compression...I didn't think the Samsung's had compression technology in them?

[Anvil]

Staying at 46% does not make a difference for the non compressing drives but it won't hurt either, it should not make a difference for speed and so he might just keep it at some none 0-Fill option.
I might go for a different option with the next drive, even if it doesn't do compression.

Interesting that the Samsung settles at about the same speed as I get on the 120GB Force 3, it has double the capacity and still the Samsung is faster.
The Synchronous version of the Force 3 (the GT or Vertex 3, equivalent) would most likely place itself closer to the Mushkin.

Are you sure that you have disabled anti-virus real-time scanning on the folder where the data is put?
(I forgot a few times myself when moving the F3 around during the bug period)

Johnw's Samsung 470 averaged ~113MiB/s so the 830 is doing fine but looks to be a little bit slower for this work-load.

[bluestang]

M225->Vertex Turbo 64GB Update:

879.17 TiB (966.66 TB) total.
~2245 hrs (Torture), 3304 hrs (Power-On).
17092 Raw Wear.
109.89 MB/s avg for the last 7.90 hours (on W7 x64).
MD5 every 20 loops on 1.59GB file = OK.
C4-Erase Failure Block Count (Reallocations) 24 to 25.
C5-Read Failure Block Count (uncorrectable bit errors) from 24 to 27.
C4 and C5 list getting to long to show.
So a total (B8+C3+C4+C5) of 62 Bad Blocks.

May let it run overnight.

Edit: I'm over 65000 loops according to the ASU log.

[Christopher]

Anvil,

The endurance testing system is as clean as a system can be, no AV scanning, indexing, or any of that stuff. I did some experimentation the first 24hrs as some combination of setup was causing the speed to drop into the 78MBs range. It's now averaging 108MBs and instantaneous estimated MiB start at 154, then slowly drops to 125. With 6 second pauses and 6 second randoms, it averages 108MBs. The less free space there is, the less the speed. It's pretty linear.

At first, it was super fast for the first few loops. Then, speed rapidly decreased. Somewhere along the line the drive stopped writing randoms and that greatly increased speed. Once I tweaked the loop parameters by increasing random write duration and the pause, I found a happy medium that preserves randoms and a respectable average speed.


I use the 46 percent compression level because I just don't like the idea of writing 0 fill to the drive. It shouldn't really matter with the 830, but most data that gets written in every day useage is somewhat compressible. Also, I have the same set-up that was run on the Mushkin, though I'm considering adding 39.3GB of static data like B.A.T.'s drives.





This might be a dumb question, but is the amount of voltage needed to write and hold 00s in a cell less than some other combination of bits in MLC?

[B.A.T]

Todays update:
Kingston V+100
283.1335 TiB
1284 hours
Avg speed 25.29 MiB/s
AD still 1.
168= 1 (SATA PHY Error Count)
P/E?
MD5 OK.
Reallocated sectors : 00


No disconnect the last 24 hours, but it's still crawling away at very low speed.

Intel X25-M G1 80GB
105,0538 TiB
19435 hours
Reallocated sectors : 00
MWI=70 to 64
MD5 =OK
49.27 MiB/s on avg


m4
13.7037 TiB
50 hours
Avg speed 81.46 MiB/s.
AD gone from 97 to 93.
P/E 234.
MD5 OK.
Reallocated sectors : 00

[Christopher]

So a few weeks ago, when the Mushkin died, I also had another SSD failure -- the Agility 60. After a lot of work, I was able to get the drive d-flashed. It just died again, and I can't seem to get it to detect with the jumper (though I had problems last time with this, and after a few hours I could detect it -- but it took about 8 flashes to work). I was running 1.7fw, d-flashed it with 1.7 a few weeks ago, and I think that could (possibly)be the problem.

So Bluestang, if you decided to Turbotize your other M225... maybe stick with 1.6FW. I've never had a problem with 1.5 or 1.6 on any of the Indys.

[bluestang]

So Bluestang, if you decided to Turbotize your other M225... maybe stick with 1.6FW. I've never had a problem with 1.5 or 1.6 on any of the Indys.

That has been Turbotized already. I did them at the same time since I was running them in RAID 0. I'll have to take a closer look at 1.7 then. Thanks!

[minpayne]

The compressibility of data only matters for SandForce drives. They're the only controllers that can compress data/deduplicate data (there may be an enterprise controller made by some other company that does so as well, but I'm not sure. I'm mostly concerned with consumer drives.)
...

Thank you for the detailed reply That did explain lots of things! (I cannot add "Thank You" onto your post yet)

Though I know it'd be luxury to destroy a large SSD, I'm interested in the endurance (and related algorithms) of the 8k-page-sized models, such like the M4 256G and 512G. It is suspected that the M4 controller is not taking good care of WL and WA and may lead to early death of certain blocks of the 8k-page-sized nand flash. I'd be interested to see how an M4 256G/512G goes compared against the M4 64G in the normalized chart.

[Christopher]

The has been Turbotized already. I did them at the same time since I was running them in RAID 0. I'll have to take a closer look at 1.7 then. Thanks!

Turbotize is the greatest word I've ever made up.

[Anvil]

Kingston SSDNow 40GB (X25-V)

536.11TB Host writes
Reallocated sectors : 05 14
Available Reserved Space : E8 99
POH 4863 hours
MD5 OK

33.83MiB/s on avg (~35 hours)

--

Corsair Force 3 120GB

01 92/50 (Raw read error rate)
05 2 (Retired Block count)
B1 47 (Wear range delta)
E6 100 (Life curve status)
E7 10 (SSD Life left)
E9 499272 (Raw writes) ->488TiB
F1 664530 (Host writes) ->649TiB

MD5 OK

106.44MiB/s on avg (~120 hours)

power on hours : 1900

WRD has been locked at 47 for the last few days, unusual!

edit:
@bluestang

I just checked my (X25-V) and it has done 37565 loops but that does not include any loops before May 31st (that's when the loops started to get recorded) and so there is missing 14-15 days of loops + a few unexpected reboots during the SF issues.

166,624,167 files has been created and deleted

[bluestang]

@bluestang

I just checked my (X25-V) and it has done 37565 loops but that does not include any loops before May 31st (that's when the loops started to get recorded) and so there is missing 14-15 days of loops + a few unexpected reboots during the SF issues.

166,624,167 files has been created and deleted

Yeah, I'm sure we are all missing some log info for various reasons...most of it out of our control.

287,950,582 files created/deleted for me

Edit:
System was totally locked up when I got in. The culprit...most likely the M225.
C5-Read Failure Block Count jumped from 27 to 37.
Most of the RF's are on Bank 5.

If the M225 makes it to 1 PB, it will do it kicking and screaming the whole rest of the way

Edit2:
@RyderOCZ...care to give me an idea on how many bad blocks the 1.6 Turbo FW has budgeted?

[bluestang]

System was totally locked up when I got in. The culprit...most likely the M225.
C5-Read Failure Block Count jumped from 27 to 37.
Most of the RF's are on Bank 5.

If the M225 makes it to 1 PB, it will do it kicking and screaming the whole rest of the way

Had to wipe disk partion, sanitary erase, full format and run CHKDSK /R to get M225 accessible to endurance test again.

Copied static data back over and ASU ran for just under an hour (.82 to be exact) doing 25 loops @ 118.84 MB/s before ASU stopped with "Write Error 0" and drive subsequently dropped out from system again.

C5-Read Failure Block Count went to 43 during this time.

[Anvil]

There is very little lost as each loop is stored (and the totals can be recalculated), still it could mean thousands of files and a few GiB per loop, I expect I could have lost up to 50GiB of writes, all in all that is nothing vs the amount of writes that has been performed.

It looks like the M225 will need a lot of care and attention to get past 1PB, let's keep our fingers crossed.

[Christopher]

Ummm, not sure this is a suitable test candidate. I'm not sure this drive is suitable for anything.

It's some random Patriot 32GB. The chassis is quite nice, but what's inside seems to be awful.

[Anvil]

What controller and NAND?