SSD Write Endurance 25nm Vs 34nm

I've updated the info in post#1 on the M3P, an extremely exciting drive, lets see how it compares to the Samsung.

@Ao1
I've been looking for data on the Kingston V100+ and it is there, I looks like we couldn't decide on whether the attribute in question really was an MWI counter, anyways I did put it in the table. TiB written when it ran out of MWI would mean ~3300 so it looks to be correct.

I've got logs for my drives from the day I started using the log, January 9th this year.
I'll find the missing "points" for the X25-V, shouldn't be that much of a job if we do our own drives.

An overview of all drives that are not running anymore.

Attachment 127140

The 460 is one of the drives that we never really decided/agreed on when MWI was exhausted.
(Johnw, looking back and vs the 830 attributes, 60TiB looks to be a very low figure for that drive)

Attachment 127142

Attachment 127141

Quote:

---

Originally Posted by Christopher

Ao1,

V4 128GB Intel 29F64G08ACME3
Octane 128GB Intel 29F64G08ACME2

---

I can’t find a data sheet for that either of this NAND types, but onethreehill has done an excellent job of collecting data on controllers and NAND.

http://forums.storagereview.com/inde...eviews-thread/

According to that link the NAND in the V3/ Octane has been used in a number of different drives. The Octane also used Hynix NAND and there have been widespread reports of premature relocated sectors with Hynix NAND. So the issue is evident on NAND from two different suppliers.

Intel 29F64G08ACME3
Kingston SSDNow V+200
Zalman F1 120GB
Team Xtreem S3

Intel 29F64G08ACME2
Corsair Force Series GT 120GB
Mach Xtreme DS Turbo 120GB
Silicon Power V30 60GB
Silicon Power V30 120GB

Quote:

---

Originally Posted by Anvil

I've got logs for my drives from the day I started using the log, January 9th this year.
I'll find the missing "points" for the X25-V, shouldn't be that much of a job if we do our own drives.

---

That would be great. There is a lot of info missing in the tables in the first post. It would be great to see a chart showing how relocated sectors occurred for each drive and also a calculation of WA/WL based on when the MWI expired (if sufficient SMART data is available).

Quote:

---

Originally Posted by Ao1

I can’t find a data sheet for that either of this NAND types, but onethreehill has done an excellent job of collecting data on controllers and NAND.

http://forums.storagereview.com/inde...eviews-thread/

---

The same info is listed in the storage section here as well.

Lol, that will teach me for only hanging around in the storage section.

The X25-M G1 used 50nm NAND and Intel claimed a write amplification factor of 1.1. The best possible is 1, so from 2008 when the X25-M G1 came out there has been near zero scope for improvement in f/w algorithms to further reduce WA.

The only way to get below 1 is to use compression. To get closer to 1 the only other option is to use DRAM to optimise how writes are placed on the NAND.

Looking at it that way it makes more sense to me now why Intel have switched to SF and why drives with no compression are having to start to use large DRAM caches to offset reduced P/E cycles (which Intel have always avoided).

It could be that drives with large DRAM caches are negatively impacted by the sustained writes that occur in the endurance workload as opposed to normal workloads that occur in real life. That does not explain the V3 or Octane however.

The 830 has a relatively large DRAM cache, and it doesn't seem to be adversely affected. I think Intel just went with SF as a hedge against smaller lithography. When they started with SF, they probably ha no idea what 20nm was going to look like in production form (with respect to longevity). At least in Intel's case, SF seems to be doing its job in that respect, but i'm not super stoked about 20/19nm regardless of what controller it's bolted too.

Assuming P/E cycles are calculated correctly with the 830, your drive hit 827,911 @ MWI 1.

Capacity *P/E spec / ((WA) * (WL)) = Total theoretical write capability

256 *5,000 = 1,280,000/ 1.54

WA & WL = ~ 1.55

I'll work it out for a drive with no cache later.

This was the post where mwi reached 1 for the Kingston V+100. Link Read error rate and reallocated sectors count was at 0. They didn't move when the ssd died.

M4v1 reached wmi=1 without a change of the Read error rate and reallocated sectors count at 172.3671 TiB. The first Read error rate came at 700.4629 TiB
M4v2 reached wmi=1 without a change of the Read error rate and reallocated sectors count at 174.3919 TiB. At 951 TiB it still have none.
Intel X25-M G1 reached wmi=1 without a change of the Read error rate and reallocated sectors count at 147,5687 TiB.. The first reallocated sector came at 308,22 TiB.

The 830 is supposed to be using 3K PE flash, but MWI was based on more like 3.6K.

Todays update:

m4
951.7772 TiB
4006 hours
Avg speed 74.35 MiB/s.
AD gone from 69 to 66.
P/E 16394.
MD5 OK.
Reallocated sectors : 00
http://www.diskusjon.no/index.php?ap...tach_id=492199

Intel X25-M G1 80GB
655.58 TiB
23318 hours
Reallocated sectors : 162
Available Reserved space: 90
MWI=124
MD5 =OK
42.79 MiB/s on avg
http://www.diskusjon.no/index.php?ap...tach_id=492203

Intel X25-E 64GB
131.75 TiB
405-30=375 hours
Reallocated sectors : 0
Available Reserved space: 100
MWI=99
MD5 =OK
101.07 MiB/s on avg
http://www.diskusjon.no/index.php?ap...tach_id=492201

Quote:

---

Originally Posted by Christopher

The 830 is supposed to be using 3K PE flash, but MWI was based on more like 3.6K.

---

There are a lot of assumptions with the Samsung due to a lack of proper SMART info and info on the NAND. The X25-V is a bit easier:

X25-V
MWI 1 @ 187,627 GB
40 *5000 = 200,000 GB
WA & WL = ~ 1.06

Here is a quick summary of drives with good SMART info. I’m going to guess that the P/E count for the Chronus was 3K not 5K giving a WA factor of 0.85, which seems more logical considering 46% compression was used.

M4 64 GB (25nm) = 176,128 – 64*3000 = 320000/ = 1.09 (or 1.8 with 5K P/E)
Intel X25-M G1 80 GB (50nm) = 150,528 – 80*10,000 = 800000/150528 = 5.31 (No TRIM)
Intel 320 40 GB 25nm = 194,560 - 40*5000 = 200,000/ 194,560 = 1.03
C300 64 GB (34nm) = 301,056 - 64*5000 = 320,000/301056 = 1.06
Chronos 60 GB (32nm) = 218,112 – 60*5000 = 300000/218112 = 1.37 (or 0.85 with 3K P/E) (46% compression)
Intel 520 60 GB = 429056– 60*5000 = 300000/429056= 0.69 (46% compression)
Intel X25-V 40GB = 187,627 – 40*5000 = 200000/187,627 = 1.06


Edit:

WA for the Intel 320 is slightly lower than the X25-V as expected, but getting any lower without compression must be close to impossible, so as 1,000s of P/E cycles disappear with every reduction in NAND geometry there is no way of avoiding a reduced life span.

Intel 330 120GB

40.72TB Host writes
Reallocated sectors : 05 0
Available Reserved Space : E8 100
MWI 97
[F1] Total LBAs Written 1334448
[F2] Total LBAs Read 22623 // no movement
[F9] Total NAND Writes 29400GB
POH 106
MD5 OK

122.48MiB/s on avg (~33 hours)

http://www.ssdaddict.com/ss/Endurance/I330_MWI.png


--

Kingston SSDNow 40GB (X25-V)

988.64TB Host writes
Reallocated sectors : 05 46
Available Reserved Space : E8 99
POH 8854
MD5 OK

34.30MiB/s on avg (~33 hours)

--

Attachment 127143

The SMART info on my 830 is fine. WLC works like it should, incrementing every 230GB, which was almost exactly 4x the 64GB Turbos. I believe it hit MWI 1 at 3620PE cycles. In any event, it's not 5K.

Assuming perfect wear leveling, PE count should increase every 256GB. It doesn't have perfect WL, so we can go by PE count (which has been the paragon of consistency this time around) 256/230 = 1.113

3600 * 256GB = 921,600

921,600 / 1.113 = 828,032GiB


I'm not really sure where I'm going with this, but ~1.1 is really where most decent drives should be (that aren't SF, but I suspect you might include them as well) with TRIM and whatnot.

The 830 is not suffering from 1.5xx WLE/WA no matter what.

The thing is that some of the firmware-enhancements may have altered and bettered both WA and WL.
Some of the drives have had the fw updated multiple times during the test, that leaves some uncertainty about how the drive could have performed if it was using the later/best fw since day 1.
Well, I'm going to update to the latest fw when/if it shows up, as would I in real life.

Anyways, it is interesting seeing what the figures look like as they should be close.

The 01FW for the 830 is a sack full of pain where WA is concerned. I think the V4s WA got markedly worse with 1.4FW. Most drives don't get that much of a bump one way or the other, and it's hard to know what's going to happen before you flash.

I know that the M3P's GC got tweaked on 1.03 FW, but I didn't see much difference. I did a lot of IOmeter and power consumption testing for several hours before puttin the M3P to work.

1.5 PB?!! That Samsung's doing great :shocked:

Congrats on becoming a mod Anvil, when did that happen?

The Torqx 2 had its first bad block at approximately 40600 GiB, with 1677 average NAND cycles, 2539 max NAND cycles.

The Vertex 4 had its first bad block at 294.27 TiB Host writes.

If you have an octane/petrol or vertex 4 and your drive gets a bad block, you better have a backup!

The Kingston (X25-V) had reallocation's (1) from the get-go.

I'm sending a list of all reallocation's to Ao1. (until SmartLog was introduced)

Quote:

---

Originally Posted by Ao1

Edit:

WA for the Intel 320 is slightly lower than the X25-V as expected, but getting any lower without compression must be close to impossible, so as 1,000s of P/E cycles disappear with every reduction in NAND geometry there is no way of avoiding a reduced life span.

---

Yes there is, multiple techniques:

a) More ECC, so the wearing effect is less apparent and takes longer to degrade the capacity of an NAND cell to hold data.
b) Gentle programming/erase techniques. Not entirely sure if this is possible, but multiple companies claim to be doing stuff like this ... it will trickle down to consumer SSDs as needed.
c) Larger capacity SSDs. As lithography shrinks, capacity of drives will increase with mostly offsets the reduction in endurance. It is nice to dream of a 512gig drive in 50nm NAND, but outside of the odd deal, no normal person can afford it.

EEC deals with correctable errors and as NAND geometry shrinks, instances of correctable errors increase exponentially. A correctable error does not necessarily mean the NAND cell is on its way out. Any improvements that may be available in the future have to be offset by the continuing growth of page sizes as geometries shrink. I really can’t see anything that is going to reduce WA below what it is already (without compression), especially if page file sizes continue to grow, but then again I’m not a NAND engineer.

Quote:

---

Originally Posted by Ao1

EEC deals with correctable errors and as NAND geometry shrinks, instances of correctable errors increase exponentially. A correctable error does not necessarily mean the NAND cell is on its way out. Any improvements that may be available in the future have to be offset by the continuing growth of page sizes as geometries shrink. I really can’t see anything that is going to reduce WA below what it is already (without compression), especially if page file sizes continue to grow, but then again I’m not a NAND engineer.

---

Error rate also increases as the NAND becomes more worn. Having stronger ECC means you can apply more wear to the NAND cells before they become too damaged to reliably correct. This is one of the keys ideas behind Intel's HET-NAND.

Bad/weak block identification also seems to be an important key to SSD long life. Intel's own controller seems to have this spot on, while I'm more skeptical of the crucial and sandforce drives. The everest (and everest 2) controllers probably need more work, something seems a bit iffy about how bad blocks occur on octane and vertex 4 drives.

I agree that WA doesn't really any room to improve anymore ... although working block dedup could certainly improve it a bit more.

I've actually moved the two drives out of the endurance rig and into my main desktop. I want to see what happens over the next week, as I noticed that the drives seem to be not in-considerably faster.

After 21 hours, the M3P is at 270MB/s and the 830 is at 303MB/s. I'm probably hallucinating, but I feel like there is something preventing the drives from running quite as fast as they should. Both rigs are on 10.6.1002, one H67 and Z68.

Perhaps I imagined it, but we'll see.

Quote:

---

Originally Posted by Christopher

Both rigs are on 10.6.1002, one H67 and Z68.

---

I am using 11.1.0.1006 on my Z77. Seems to be working pretty well.

Could power management be a factor.

Maybe you are wearing out the SATA cables with all the transferred data :D