SSD Write Endurance 25nm Vs 34nm

[Christopher]

Samsung 830 64GB Update, Day 49
FW:CXM02B1Q

TiB Written:
318421.52

GiB written:
310.9585

Avg MB/s
132.51 up from 129.20

PE Cycles
15300 up from 15107

Reallocated Sectors
40960
20 Blocks, holding steady

1163 hours




OCZ Vertex Turbo 64 Update Day 5

TiB 20.4309
GiB 20921.24
Avg MB/s 83.03

Avg Erase Count
496 up from 370

MWI 91 down from 93


1 Program Fail
19 Erase Fail (up from 16)
0 Read Fail
105 Hours

[bluestang]

There is no way that is an old school Indilinx Barefoot Turbo with 51nm Samsung. And if it is, then OCZ did something wacky to supply it with FW 1.7.

[Christopher]

There is no way that is an old school Indilinx Barefoot Turbo with 51nm Samsung. And if it is, then OCZ did something wacky to supply it with FW 1.7.

Bluestang...

It's okay - you can say it out loud. You are among friends here, so do not hesitate to accuse them of doing wrong.

I think the drive --and the 128GB I bought a few months back-- could be REFURBS.

Now, I'm not really mad if they are -- they should have a 3 year warranty (which is going to be hilarious if I have to RMA them in 2015). But if they're refurbs, they should tell a homie first. I did ask one well known OCZ person by email about whether it could be truly possible to get 51nm Samsung NAND on a new drive with new FW and the new plastic chassis... and they gave me a most puzzling reply. So took that to mean something shady was going on, but I'm just happy to have them.

I would expect OCZ to have some Samsung NAND drives laying around for RMAs... but I doubt they're doing any production runs. I bought a SF powered Vertex LE 100 and the Turbo 128 from the same etailer. The 64GB Turbo was from elsewhere, but it was new and wrapped. All 3 drives had new OCZ plastic chassis. They were all sold as new.

On one hand, you have the fact that if OCZ had some "almost new" drives laying around, they could just factory flash them and put them in a new plastic chassis. On the other hand, that seems like a lot of work for such a small number of drives. If you were to rma an old Turbo, the one you would receive in it's place would most likely be in a plastic chassis.



I also bought a Vertex EX, but it was in the metal case with 1.3FW (Incidentally, the Vertex EX on 1.7FW is a beautiful thing).

[Ao1]

Host writes: 14,708 GiB
MWI: 16
P/E Cycles: 3,039
POH: 401
Relocated sectors: 8,192
MB/s: 5.97

[bluestang]

Bluestang...

It's okay - you can say it out loud. You are among friends here, so do not hesitate to accuse them of doing wrong.

I'm not accusing, just stating a fact.

Also, don't think that it will make it very far. Already 19 Erase Failure and 1 Program Failure block counts.
But, I hope I'm wrong on that so you can have some fun torturing it.

[Christopher]

I'm not accusing, just stating a fact.

Also, don't think that it will make it very far. Already 19 Erase Failure and 1 Program Failure block counts.
But, I hope I'm wrong on that so you can have some fun torturing it.

I think I would have fun torturing it regardless...

The Samsung wrote 11170.0GB yesterday, which is probably a new record for raw writes writen in a day. It's also popped another block... finally.

The WA is now normal for a modern drive, but the initial WA on the drive for the first month was enough to put a lot of wear on the drive, which is why I thought it strange that no new bad blocks had been detected.

[devsk]

That 830 seems to be in a real hurry....

[Anvil]

Lets see if it can outlast the Samsung 470.

There's about 170TiB to go...
Im not sure about the WA on the 830, it could just be that the reporting was wrong but it does look to have changed since the fw update.
(a pity it wasn't released a week or two earlier)

[Christopher]

So I just received the 16GB MOBI, which I was planning on testing as the first SLC drive.



I''m still on the fence about how long it might take and whether there was any point to testing it.
There is no smart data to speak of.

[Anvil]

It looks a bit slow but I suggest you give it a few hours on the Endurance test so we can see what it can do.

16GB isn't much so if you used 50% static and set min free space to 1GB it might do alright.

What SMART data is available?

[Christopher]

After 40min of loops, avg mb/s is sitting at 51.6MBs. I've got no static data on it, and have 2 GB min free space.

I don't know what the raw value is. It is not sectors written.

[Anvil]

If you stop the test and do some writes e.g 10GB you should be able to figure out what the raw value represents.

It's a great performer on reads (considering it's capacity/age) and I consider 50+MB/s to be an excellent result for such a drive.
It's not really that interesting (as it's old) but it would of course be interesting to see what an SLC can do.
Does it get hot?

[Anvil]

Kingston SSDNow 40GB (X25-V)

656.36TB Host writes
Reallocated sectors : 05 21
Available Reserved Space : E8 99
POH 5921
MD5 OK

33.77MiB/s on avg (~51 hours)

--

Corsair Force 3 120GB

01 87/50 (Raw read error rate)
05 2 (Retired Block count)
B1 44 (Wear range delta)
B6 1 (Erase Fail Count)
E6 100 (Life curve status)
E7 10 (SSD Life left)
E9 764525 (Raw writes) ->747TiB
F1 1017382 (Host writes) ->994TiB (~30TiB left until 1PiB is reached)

MD5 OK

106.72MiB/s on avg (~51 hours)

power on hours : 2894

--

[Christopher]

The drive does not get hot, but it's in a plastic chassis and not metal.

50+MBs is fantastic for a 16GB drive. The 16GB 3.5" 7000 PRO I have is not much faster, but I'd have a hard time cramming it in the testing rig, so I ordered this one.

I copied 10,500K to the mobi, and the raw value went from 435966 to 436009, an increase of 43.

The number doesn't seem to go up consistently.

[adrianu]

Guys, do drives, that "require" over-provisioning (like 830), need free space unpartitioned, or any free space (partitioned already) will do for GC to work normally?

[Christopher]

Guys, do drives, that "require" over-provisioning (like 830), need free space unpartitioned, or any free space (partitioned already) will do for GC to work normally?

I don't think the 830 requires over-provisioning, but I do think it's a good idea. Drives with TRIM work with free partitioned space and spare area. Drives which rely on GC (and do not have trim)can really use the extra unpartitioned space.

[Anvil]

256K chunks is not unlikely, it might be the actual block size for that NAND?

@adrianu

If the 830 is running with TRIM enabled you should be OK without reserving extra spare area. (over-provisioning)
It depends on what the drive is used for, normal usage you should be fine using the full capacity whether TRIM is enabled or not.

Without TRIM I'd personally set user-size to 110GB or thereabouts but it isn't required of course.

[Christopher]

Well, I spoke too soon. I can't really figure out what the relationship is with the MOBIs smart value to writes. It could be some kind of bit error counter.

[Ao1]

Lets see if it can outlast the Samsung 470.

There's about 170TiB to go...
Im not sure about the WA on the 830, it could just be that the reporting was wrong but it does look to have changed since the fw update.
(a pity it wasn't released a week or two earlier)

It is rather annoying that two f/w updates got issued within days of starting the test. Still I have separate log files for each f/w version, which I have combined together for the graphs below.

Prior to the 1st f/w update the #177 steps due to the fact that it did not update without a reboot. The subsequent big jump in the total LBA’s was due to a temporary switch over from the 4K endurance load to the normal endurance load.

*The 177 value is to the power of 10 to improve the scale visibility.



Now I add some trend lines. It seems that the wear has been linear regardless of workload or f/w version. I thought the enterprise workload would kill the drive in no time, but in practice it has not degraded the drive any faster than the normal endurance workload. The big difference of course being the write speeds.

[Anvil]

@Ao1

That is interesting, I expect the large cache plays an important role in relation to that the workloads results in the same wear.
So, even if the OS can't do write combining the drive is capable of doing so, I find that impressive as the test is full span.
I don't think a higher QD/Thread count would have changed anything but it would of course have to be tested.

@Christopher
There is a bit of interesting info on the Mobi out there Link

7.2.4 Total erase count
Total Erase Count means the summation of erase counts of all blocks in SSD.
In general, the maximum value of 32GB SSD's Total Erase Count is 3200M and if the Total Erase Count exceeds the
value, it says that the SSD is completely wore - out.
Maximum Total Erase Count differs from vendors of NAND FLASH Chip and the capacity of SSD.

7.2.5 Total erase count current value
This value becomes 100 when Total Erase Count is equal to 0 and becomes equal to 0 when Total Erase Count is
maximum value.
If the value is equal to 0, it means that the SSD is completely wore-out.

7.2.6 Total erase count worst value
This value becomes 100 when Total Erase Count is equal to 0 and becomes 0 when total erase count is 90% of maximum
value.
Users should be warned when this value becomes equal to 0.

7.2.7 Total erase count raw value
This value means the summation of erase counts of all blocks in SSD.

You'll find it on page 17 and on-wards.

So, the current value acts as MWI

[Christopher]

Over the course of the test, the 830 has used 1 PE cycle per 23GB. But if you look at just the last two days, that number is 58GB/PE cycle.

If the drive died right now, it would be a decent result in terms of PE cycles. But if the numbers are accurate, and it makes it pretty far, it will be a good story.

The Vertex Turbo is using 1 PE cycle for every 56GBs.

[Christopher]

Samsung 830 64GB Update, Day 50
FW:CXM02B1Q

GiB Written:
329588.84

TiB written:
321.8641

Avg MB/s
132.36

PE Cycles
15494 up from 15300 <--- this is not a lot, it used to do 370+ a day.

Reallocated Sectors
45056
22 Blocks, up from 20!

1187 hours




OCZ Vertex Turbo 64 Update Day 6

TiB 35062.37
GiB 34.2406
Avg MB/s 83.09

Avg Erase Count
621 up from 496

MWI 88 down from 91


1 Program Fail
19 Erase Fail (up from 16)
0 Read Fail
129 Hours

[Ao1]

Host writes: 15,236 GiB
MWI: 11
P/E Cycles: 3,232
POH: 401
Relocated sectors: 8,192
MB/s: 5.95

[Ao1]

Here is a graph showing writes per day and changes in the P/E count. Once the P/E cycles updated correctly the drop in the P/E count is linear to the write activity.

[Christopher]

I need to run some benchmarks on the 830 -- While average writes are fast, and getting faster, I still don't know if reads are back to where they should be.

It's just that, damn... the 830 is fast right now. Really fast. It's creeping up on 134MB/s avg (and instantaneous readings throughout the loop are in the 143 - 147MBs range). That's faster than the Mushkin Chronos D at 46 percent compressibility.