SSD Write Endurance 25nm Vs 34nm

[Vapor]

Updated charts

C300 isn't in the MWI Exhaustion graphs and axes haven't been adjusted in the Writes vs. Wear graphs because it's just too soon...MWI is down to just 99, very hard to extrapolate with accuracy off of that.

Host Writes So Far


(bars with a border = testing stopped/completed)


Raw data graphs

Writes vs. Wear:


MWI Exhaustion:



Normalized data graphs
The SSDs are not all the same size, these charts normalize for 25GiB of onboard NAND.

Writes vs. Wear:


MWI Exhaustion:



Write-days data graphs
Not all SSDs write at the same speed, these charts factor out write speeds and look at endurance as a function of time.

Writes vs. Wear:


MWI Exhaustion:

[One_Hertz]

179TB. 7%. 17 Reallocated sectors.

[bulanula]

Vapor, even the factory bad block count seems to have changed?

John's Samsung test is over now?

Hertz getting close to 0 mwi now

[Vapor]

Factory bad block has always been 98. First screenshot was in hex and 62h = 98.

johnw is at MWI = 0, but he can keep testing until it dies or until read errors occur (via Anvil's MD5 update). And yeah, One Hertz is the next to get to MWI = 0.

[Ao1]

MWI 1 (0 does not exist on Intel drives)

[Vapor]

Playing around with F8, trying to figure out if it is NAND writes....can't figure out units if it is though

Take these two data points:

4.3924TiB, 26072470
1.0822TiB, 4446418
________________-_
3.3102TiB, 21,626,052....F8 increased by a factor of 5.86x; writes increased by a factor of 4.06x.

The 5.864:4.06 ratio (1.44x) seems too much like a good candidate for write amplification for it to not be investigated, IMO.

3.3102TiB = 3,389.6GiB = 3,470,996 MiB = 6,942,000 blocks (512KiB) = 888,575,000 pages (4KiB) = 3,554,300,000 KiB

So each unit of F8 could be equal to: ~164KiB, ~41.1 pages, or ~.321 blocks.

Each unit of F8 divided by 1.44 could be equal to: ~236 KiB, ~59.2 pages, ~.443 blocks.

None of it works out though

I suppose I'll also need to pause Anvil's app for awhile and see how things move.

EDIT: what is the block size of 34nm NAND? I've been assuming 512KiB, but that may not be the case it seems...

[One_Hertz]

I always thought the block size of 34nm NAND was 128 x 4KiB = 512KiB.

[Ao1]

I found a bit of info here












Attachment 117318

[Eggcake]

Great thread, very nice work

Biggest findings so far (for me at least) :

1. All SSDs (even the Samsung 470) endure pretty heavy writes - I doubt I will ever write that much on my SSD...
2. 25nm doesn't harm the Intel drives (regarding lifetime)
3. SandForce is a huge disappointment. I expected a drop in performance...but seriously, the SDcard in my smartphone writes faster than this...that is unacceptable IMHO - but I don't want to kick off another discussion about this, as there is already another thread. I just didn't expect this at all.

I'm looking forward to more results

[B.A.T]

129 hours, 37,3114 TiB, Wear Leveling Count and Percentage of the rated lifetime used has gone from 82 to 79.
Avg speed for all 129 hours is roughly 84,2 MiB

[bulanula]

Nice findings ao1

[Vapor]

6.38TiB, 98MWI, 106 raw wear, 29hrs

[johnw]

62.535 TiB, 178 hours, sa177: 1/1/5230

65.521 TiB, 189 hours, sa177: 1/1/5483

Note that there were a few power-on hours when Anvil's app was not running between last update and this one. Therefore the difference between these last two updates should NOT be used when calculating average write speed.

Anvil's app is now reporting average write speed. At the moment it is saying 106.84 MB/s. I was getting about 110 MB/s, but now the app is doing a 2.1GB file copy and MD5 sum check every 10 iterations, so I think that is slowing things down slightly. I hope to set it to every 100 iterations once the app allows it (in the next app update).

[devsk]

65.521 TiB, 189 hours, sa177: 1/1/5483

Note that there were a few power-on hours when Anvil's app was not running between last update and this one. Therefore the difference between these last two updates should NOT be used when calculating average write speed.

Anvil's app is now reporting average write speed. At the moment it is saying 106.84 MB/s. I was getting about 110 MB/s, but now the app is doing a 2.1GB file copy and MD5 sum check every 10 iterations, so I think that is slowing things down slightly. I hope to set it to every 100 iterations once the app allows it (in the next app update).

That Samsung is putting up hell of a fight...

[B.A.T]

139,5 hours, 40,4361 TiB, Wear Leveling Count and Percentage of the rated lifetime used has gone from 79 to 77.
Avg speed for all 129 hours is roughly 84,4 MiB

[Anvil]

138.60TB Host writes
MWI 24
MD5 OK

No other changes.

[Computurd]

now all we need is a V3. would be interesting, very very interesting, to see what form throttling takes on the new gen of SF drives.

[Anvil]

Yes, some more SF and the SF2 series in particular and some sort of HDD would be interesting as well

[Ao1]

The SF2xxx drives are also lifetime write throttled. It will not allow PE cycles to exceed the warranty life time.

The only way to run an accelerated wear out test on a SF drive is to find one that has not had the throttling option set by a vendor.

[bluestang]

^^ Come on Zads...I know you can help out here!

[Anvil]

The SF2xxx drives are also lifetime write throttled. It will not allow PE cycles to exceed the warranty life time.

The only way to run an accelerated wear out test on a SF drive is to find one that has not had the throttling option set by a vendor.

We might have some other brand in line for the test, could be the same throttling though.

[One_Hertz]

Really need a way to remove throttling on SF first. If only I knew the parameters to specify for their FORMAT command.

183TB. 5%. 18 reallocated sectors.

[Ao1]

SMART attribute 181 (B5)

Below is some clarification on non aligned reads/ writes:

"The NAND page size is handled in firmware, before the SMART calculation is performed. Firmware has internal counters for total unaligned writes and total unaligned reads (regardless of page size). These counters are added together, divided by 60,000 to get the raw value.

It should be noted that this number is not related to sector size, but rather to NAND page size (though I suppose that depends on how you define “sector”). The counters are incremented in one of two cases: 1) the write starts somewhere besides a page boundary, or 2) the write from the ATA command is not the length of the physical page (be it 4k or 8k). The math is the same, regardless of page size. This is because for the 8k NAND on the 256GB and 512GB, we simply take 2 4k chunks from the host at a time, and write them to the 8k page. If they are misaligned for 4k, they’ll be misaligned to the same degree for 8k."

[Hopalong X]

On the Intel 320 you can track what appears to be a reverse MWI if someone knows how to access these logs listed in the Snip.
I have no idea if this can be done on the X25.
Just something I ran across.

[B.A.T]

151 hours, 43,9010 TiB, Wear Leveling Count and Percentage of the rated lifetime used has gone from 77 to 75.
Avg speed for all 151 hours is roughly 84,7 MiB

There're no indication of the speed decreasing. The M4 looks very good so far and with 3/4 of the journey left there is no sign of reallocated sectors.