SSD Write Endurance 25nm Vs 34nm

[devsk]

I think the M4 will die soonish now.

Define soonish and let's take bets...

[Computurd]

That M4 isnt going anywhere for a long time.

[bulanula]

Maybe upgrade the FW to be as realistic as possible on the M4 ?

[sergiu]

I think the M4 will die soonish now.

Yes... 26 september is the due date... But don't worry, M4 behaves just like a cat so it has 9 lives. It will recover itself and will become even stronger (pecentage of rated life will be resetted again to 255).

[bluestang]

Maybe upgrade the FW to be as realistic as possible on the M4 ?

He already tried, that's where he picked up the extra attributes.....pay attention

[bluestang]

M225->Vertex Turbo 64GB Update:

310.11 TiB (340.97 TB) total
914.08 hours
6730 Raw Wear
112.70 MB/s avg for the last 16.50 hours (on W7 x64)
MD5 OK
C4-Erase Failure Block Count (Realloc Sectors) at 4
(Bank 6/Block 2406; Bank 3/Block 3925; Bank 0/Block 1766; Bank 0/Block 829)

[bulanula]

I noticed he tried but maybe try again. I think the controller is very dodgy if new attributes come up without flashing new FW just by trying to flash it etc. and also why can you not go from 0009 to 0002 etc. and downgrade etc. Seems very strange to me.

[Christopher]

Sorry to leave you guys hanging - that was kind mean of me, but I figured I'd get the shipping notice on my trip. I went on a trip to the Smokey Mountains for a few days, and just got back. But I did get shipping confirmation today that my testing drive shipped.

It's a Sandforce 2281 controlled Mushkin Chronos Deluxe, but in the super spiffy 60GB capacity (it just came out this week I guess, because it wasn't even listed in the 60GB capacity at Mushkin). It's equipped with 32nm Toshiba Syncronous toggle NAND and should be pretty fast. While I was gone, it showed up on the Mushkin website, which also sells direct. I was afraid that some of the sites were incorrect about the product listing or the availibility, because I wasn't aware that anyone was making these.

http://mushkin.com/Digital-Storage/S...CR60GB-DX.aspx

Mine shipped, but won't be here until Tuesday. I was hoping I'd have it on my door step when I got back, but my 3G cellular modem doesn't work way up there.

I wouldn't have even known to look for it, but I stumbled upon a website called FutureStorage.co.uk which sells and possibly rebrands/orders drives to their spec. I saw that they had distribution and manufacturing in North America, from a factory in Texas. I believe this factory also makes the Mushkin drives as well, so I looked to see if they made a Toggle nand SF in the 60GB capacity like FutureStorage does (which they did,). I thought a Toshiba equipped Sandforce drive would stand up really well with the drop down to 60GB.

Here's hoping it doesn't have LTT.

[B.A.T]

Maybe upgrade the FW to be as realistic as possible on the M4 ?

I noticed he tried but maybe try again. I think the controller is very dodgy if new attributes come up without flashing new FW just by trying to flash it etc. and also why can you not go from 0009 to 0002 etc. and downgrade etc. Seems very strange to me.

I tried this but no matter what I trick I used it just didn't want to update to 0009.
More bad news from me. AVG forced a restart of the rig today and now I can't reach it with teamviewer so we all need to wait for an update until 26th.
It's just the worst kind of luck that the update came the day after I'm gone for 10 days.....

Because the rig was restartet ASU is not running so we will continue from my last results I guess.

[Christopher]

Speaking of the X25-E... here's my 32GB one with the most host writes. MWI of 90 still at 580TB written:

I've noticed that one of my Intel drives jumped from 200 power on hours to 400... in about 3 hours. Luckily, the X25-E I got on Ebay only had 288GB of host writes... not that it matters. 600TB is really only like 2% of what an the drive could probably write if you look at Intel's spec vs. reality. I think my X25-Vs will last for 700,000GB of host writes, vs whatever Intel rated them for. I really, really think Intel low-balls PE cycles for all of their drives. For instance, I read a short article about Intel's forthcoming HET-MLCs...

http://techreport.com/discussions.x/21644

This is laughable. The author (I respect this site by the way) states that Intel claims a 320 series, 300GB drive is capable of writing a massive... 30TB in its lifetime? And that with 30x the endurance, 710 series 300GB drives should be capable of writing 1.1PB...

It's good for a laugh.

[Computurd]

think of what the 710 could really do...jesus 30X (projected) of the results we have seen with the 320 would be simply insane...not to mention that the 320 is still running.

Although Intel doesn't set aside any flash capacity as spare area to start, the 100GB drive does have more than 100GB worth of NAND chips onboard. The extra NAND capacity (Intel won't say exactly how much) is dedicated to a RAID-like redundancy scheme that calculates parity bits to protect against data loss due to unexpected flash failures.

very thought provoking that they could achieve this type of performance/endurance with no over-provisioning dedicated to spare area, only as XOR.

also no mention of 25nm SLC anywhere even this "late" in the 25nm game...

[Christopher]

think of what the 710 could really do...jesus 30X (projected) of the results we have seen with the 320 would be simply insane...not to mention that the 320 is still running.



very thought provoking that they could achieve this type of performance/endurance with no over-provisioning dedicated to spare area, only as XOR.

also no mention of 25nm SLC anywhere even this "late" in the 25nm game...

Also at the same site, in a different article, is a discussion of Larsson Creek's successor which utilizes 25nm SLC.

http://techreport.com/discussions.x/21652

Ask and ye shall receive (some random PowerPoint slides from IDC).


What I found funny about the HET-MLC article was actually the fact that Intel states the write span of a 300GB 320 as only 30TB -- only 100 times it's capacity. That's like saying an X25-V can only write 4TB over it's life span. Talk about under-promising and over-delivering.

I can't wait to tear into this Mushkin. I'm not sure what SF does with 32nm toggle nand as far as sacrificing one whole die... I thought they were doing something different for the 60GB capacities. Maybe RAISE isn't enabled for the 60GBs? If it works as well as Intel's system (and is enabled on the Chronos Deluxe 60, then it could last...... a lot further than this "premium Toshiba Toggle NAND's" PE cycles would indicate. Not to mention, if your workload is at all dedupe friendly, you get sub 1 WA.

The Chronos Deluxe will be my first SandForce drive (all of my drives are Intel, Indilinx, or Micron controlled), but honestly, I'm pretty impressed with the new Vertex Turbo that came while I was away. I played with it today and it's pretty damn fast for a two year old drive with 50nm Samsung NAND, but the SF2281 controller could be a whole other level, even on the Intel 510, in terms of both performance and longevity with 32nm flash. It's assembled in the USA like OWC's Mercury Pro's. I've already set up my desktop for silent running and ran ASU on one of the Agilities for 8hs in preparation.
I can't wait.

[Computurd]

Ask and ye shall receive

wow! first mention of 25nm SLC ive seen yet havent had time to dig through IDF slides yet, guess i should give that a whirl

[Christopher]

wow! first mention of 25nm SLC ive seen yet havent had time to dig through IDF slides yet, guess i should give that a whirl

That was the only one.

The next iteration of Larson will probably be around the same size with the same specs cause I don't see Intel making it cheaper than $120. I don't see them making it much more expensive though considering its supposed to be paired with SRT. They'll somehow manage to make it unattractive in relation to capacity and specs as to not steal their higher margin enterprise market share. If they could make it like the 32gb Es for $120, no one would buy anything else. I don't really believe in the HET stuff anyway, so at least a couple people would think an inexpensive SLC with TRIM is a good option. So that's my rational for not getting too excited about the next generation of 'Creeks. Hopefully they'll prove me wrong.

[Ao1]

What I found funny about the HET-MLC article was actually the fact that Intel states the write span of a 300GB 320 as only 30TB -- only 100 times it's capacity. That's like saying an X25-V can only write 4TB over it's life span. Talk about under-promising and over-delivering.

Those specs from Intel are most likely based on 4K random, full span, which is an entirely different scenario to what is being tested here.

@ Anvil: how about a version of the endurance test that runs 4K random, full span, just to give a comparison to what is being tested here? I’d be up for testing another SSD in that scenario.

Regarding the 710; spending $6.50/per GB when the only notable benefit is endurance is going to be a tough call regardless of application. Something that lasts that long, at that price, when SSD technology is still evolving doesn’t seem to make any sense at all.

[Ao1]

I can't wait to tear into this Mushkin. I'm not sure what SF does with 32nm toggle nand as far as sacrificing one whole die... I thought they were doing something different for the 60GB capacities. Maybe RAISE isn't enabled for the 60GBs?

OCZ have confirmed that their 60GB drives do not have RAISE. Most likely it will be the same with Mushkin. It will be interesting to see if Mushkin have enabled LTT. What compression level will you run Anvils' app at?

[bulanula]

Those specs from Intel are most likely based on 4K random, full span, which is an entirely different scenario to what is being tested here.

@ Anvil: how about a version of the endurance test that runs 4K random, full span, just to give a comparison to what is being tested here? I’d be up for testing another SSD in that scenario.

Regarding the 710; spending $6.50/per GB when the only notable benefit is endurance is going to be a tough call regardless of application. Something that lasts that long, at that price, when SSD technology is still evolving doesn’t seem to make any sense at all.

That is what I was thinking all along. Companies are not stupid. They would not want to underestimate their products so badly. What we are doing here is just dumping a load of sequential data onto the SSD in a precise pattern. 4K random is totally different and much more stressing. As is secure erasing, which nobody thought would be good to test etc.

Just put one of these MLC 25nm drives like the Intel 320 into an enterprise server environment and see how long it really lasts. I bet it would be pretty close to the 30TB Intel specify it to last etc. and NOT 300TB that our tests have shown otherwise why would anyone buy SLC drives. Or think about cache drives. Why would Intel use SLC 34nm in the Larsen Creek if MLC was so good as we have shown here in these tests ? Does not make sense to me. Maybe I just don't get it but I think Ao1 is thinking the same thing.

Sure, our test is good for the "average" user at home that stores static data and mostly transfers sequentially but this cannot compare to enterprise usage with no static data and where all the drive capacity is used 24/7 for caching or similar 4K random intensive tasks. Hence, Intel's rating difference to what we saw here.

[Anvil]

..
@ Anvil: how about a version of the endurance test that runs 4K random, full span, just to give a comparison to what is being tested here? I’d be up for testing another SSD in that scenario.
..

Consider it done, will have a look at it today, if it was to be close to "enterprise" workloads I'd say we need to agree on some QD.
I'll probably make QD configurable but in general we'd need to agree on some value.

edit:
@bulanula
Keep in mind that the drives we are testing are not Enterprise drives although Intel hints on OP-ing their drives sort of makes them usable in "business scenario's"

[Anvil]

324TB Host writes
Reallocated sectors : 7 (we finally have some movement here and it happened during the last 68 hours )
MD5 OK

Power on hours is now at 2995 and it has been writing continuously for 2860 hours.

Will have a look at random writes shortly, should be some TBs.

edit:
1,07TiB 4K random writes (1125478MiB)

I had expected more but this drive is not a great performer at random writes, would be interesting knowing what the 320 has generated (One_Hertz could you check what is logged ), I expect it could be 2x or more.

[Anvil]

I'm just about to start the Endurance test using a Force 3 120GB.

I've been preparing a bit so it's got ~6hour runtime so far, mostly endurance related.
I'll do some final test on the rig where the other drive is running and if all is well I'll start in a few hours.

ASU_START.PNG

START_CDI.PNG START_SSDLIFE_FREE.PNG

[Ao1]

Below are the Intel specs for enterprise. For client it is around 35TB (20GB per day *5 years)

@Anvil QD is tricky. As far as I can see Intel don’t mention the QD they use for their endurance specs.

[devsk]

Intel publishes lower numbers for 320 series because they want to sell 710 series to the enterprises at a much higher price. Look how good the table above makes 710 series look. Just put the number 500TB in the first table for 120GB drive and nobody will touch the 710 drive.

I wonder why nobody notices that after having seen nearly 1700 posts in this thread. And this thread just talks about MLC consumer grade drives!

@bulanula: There is no true 100% random 4KB write enterprise scenario. There are always sequential writes to ease the drive off. Enterprise filesystems (ZFS) as well as databases have tricks up their sleeves to convert some of those random 4KB writes into larger sequential writes. Note "some".

[bulanula]

edit:
1,07TiB 4K random writes (1125478MiB)

Confirms that this testing is not what would be deemed "enterprise usage". If this value would have been closer to 30TB then the drive would surely be dead by now. Really interested to see testing using the new fully 4K random methods !

[Computurd]

@Anvil QD is tricky. As far as I can see Intel don’t mention the QD they use for their endurance specs.

ive seen 32 in a few of their documents, i would need to dig them up though.

EDIT: that was with the x-25m, but im sure it is one in the same.

[sergiu]

Confirms that this testing is not what would be deemed "enterprise usage". If this value would have been closer to 30TB then the drive would surely be dead by now. Really interested to see testing using the new fully 4K random methods !

Enterprise applications are build to sustain the load while are kept on fast RAID HDD arrays which can barely provide 10-20MiB/s of random read/write. Now, usually when HDDs are replaced with SSDs, the next bottleneck becomes the CPU as data is fed very fast. Also most enterprise applications are optimized for reading/writing big chunks of data of at least 0.5-1MB (on a hdd, due to it's rotational latency, it takes around the same time to read/write sequential chunks with size between 4KiB and 1MiB). Making enterprise applications that are totally dependent on 4K random read/write is purely retarded. And as an example, Google is using a file system for their clusters which is optimized for files with an average size of 10-60MiB. Their operations are sequential in nature, as this is the best way to use the hardware. Even database engines are highly optimized as MySQL can use 64KiB chunks for read/writes.
Now, the worst possible scenario is indeed using a SSD as a cache, but also, keep in mind that the usage model for a cache is random writes + random reads where depending on the scenario, you might have a higher number of reads than writes which will decrease the average writing speed significantly to allow a recovery period for NAND cells. If real enterprise usage test usage is desired, then the SSDs should be tested in such a way that each page that was written should be also read at least once. Pure random writes on the entire space would be torture, not real life usage.