SSD Write Endurance 25nm Vs 34nm

[johnw]

Samsung looks like the seabiscuit of this race.... it refuses to give up!

Heh, but I still would not bet on the Samsung to write more than either Intel before failure. Even though the Samsung has 64GiB of flash on board and the Intels have only 40GiB and 48GiB, I suspect the Samsung has a much higher write amplification, which is a significant handicap. But I am not certain about the WA, so I could be wrong about that. We will see!

[johnw]

1000TB is a lot, but maybe the intel 320 will go so far. It's got 16GB extra nand compared to the rest to handle less P/E .

A 40GB Intel 320 has six 8GiB flash packages on its circuit board. I think that five of them are used normally (five channels of the controller), and the sixth is used for the RAID-4 like XOR parity. So that would be 8GiB extra flash.

[Hopalong X]

Has anyone heard from One Hertz? It has been a few days since he posted or I missed something.

[Vapor]

A 40GB Intel 320 has six 8GiB flash packages on its circuit board. I think that five of them are used normally (five channels of the controller), and the sixth is used for the RAID-4 like XOR parity. So that would be 8GiB extra flash.

The Intel 320 40GB has 48GiB of NAND? Hmmm, that would put its WA at like 1.2x, which I find odd (being worse than the X25-V). Could that have been as a trade for speed compared to the X25-V? Does it really need 28.8% (37.27GiB usable on 48GiB NAND) spare area?

Either way, 48GiB of NAND changes the charts.

C300 Update:

62.53TiB, 79 MWI, 1052 raw wear indicator, 61.85MiB/sec, MD5 OK

[Anvil]

I'm not sure about the NAND on the 320 40GB, I'm sure One_Hertz can find out, I'll check mine as well as soon as I can get that torx screwdriver. (will try tomorrow)

I'll make my next report in about 8-9 hours, been away for the weekend and the rig has been powered off.

[johnw]

The Intel 320 40GB has 48GiB of NAND? Hmmm, that would put its WA at like 1.2x, which I find odd (being worse than the X25-V). Could that have been as a trade for speed compared to the X25-V? Does it really need 28.8% (37.27GiB usable on 48GiB NAND) spare area?

No, no, no!

The 40GB 320 does indeed have 48GiB of flash on its circuit board, that is a fact (no need for anyone to double-check unless you really want to), but you need to read what I wrote! Only 40GiB of the flash is used for normal operation. The extra 8GiB is used for XOR parity data. So WA should be calculated assuming 40GiB of flash.

That is similar to Sandforce SSDs where, for example, a 120GB Sandforce drive usually has 128GiB of flash on-board, but only 120GiB of the flash is used for normal operation (data storage and reserved space), while the extra 8GiB is used for so-called RAISE, which is Sandforce's name for RAID-4 like parity. Although I am less certain with Sandforce (as compared to Intel 320) how they actually implement RAISE. It could be RAID-5 like, as far as I know. But I'd guess RAID-4 like.

[johnw]

I'm not sure about the NAND on the 320 40GB, I'm sure One_Hertz can find out, I'll check mine as well as soon as I can get that torx screwdriver. (will try tomorrow)

Your SSD has Torx screws? All of my Intel SSDs use phillips-head screws. Anyway, a 40GB Intel G2 SSD has five flash packages on-board, 64Gibit each. The 40GB Intel 320 has six.

[One_Hertz]

217TiB. 24 Reallocated sectors. Anvil - I am trying to determine if I can locate that hidden smart variable. I got the output from your app and will start tracking it to see changes.

[Dragoonseal]

No, no, no!

The 40GB 320 does indeed have 48GiB of flash on its circuit board, that is a fact (no need for anyone to double-check unless you really want to), but you need to read what I wrote! Only 40GiB of the flash is used for normal operation. The extra 8GiB is used for XOR parity data. So WA should be calculated assuming 40GiB of flash.

That is similar to Sandforce SSDs where, for example, a 120GB Sandforce drive usually has 128GiB of flash on-board, but only 120GiB of the flash is used for normal operation (data storage and reserved space), while the extra 8GiB is used for so-called RAISE, which is Sandforce's name for RAID-4 like parity. Although I am less certain with Sandforce (as compared to Intel 320) how they actually implement RAISE. It could be RAID-5 like, as far as I know. But I'd guess RAID-4 like.

They don't specify on their current slides for RAISE, but their older ones that they provided to Anandtech said "like RAID-5." They wouldn't go into detail about it, however.

[Computurd]

It is amazing to me that they could dedicate such resources to parity on these devices. I wonder at the sophistication of the parity scheme. For instance, if you look at many raid controllers, etc (same concept essentially as ssd with its controller and nand) running a parity raid set can really cripple write speed in many scenarios. And these full blown raid controllers are with chips that are immeasurably faster than the low wattage chips present on an SSD.

[Vapor]

I knew the 160GB and larger versions had parity, kind of figured it was dropped from the smaller sizes for $$$$ reasons. I never knew this about the smaller 320s, very interesting. Went and read more on the 320, seems the parity was introduced to make up for 25nm deficiencies (inline with what SandForce/vendors did when going to 25nm, just doing the opposite of taking away usable space and calling it the same size on the label).

Frankly, I would prefer to calculate WA based on total onboard NAND regardless of its designed usage (WA inflated from parity doesn't seem so bad as long as it's explained...after all, for every 1 byte sent, ~1.2 bytes do get written). But we don't know exactly how much parity data is being written so I won't (not sure it's a safe assumption that the full 8GiB of the sixth die is used for parity). I'll continue backwards calculating WA using wear indicators multiplied by total non-parity NAND (not that WA is fluctuating for any of the drives). And with the Sandforces, no need to backwards calculate WA because of SMART 233.

As for normalized writes, it's probably easiest to base it on IDEMA capacity. This means everything, relative to the other drives, is unchanged except the 40GB V2, which I had based on 48GiB and will now base on 40GB. This means 55GB for the '60GB' 25nm SF-1200 drive (not sure what SF-2200 has for capacity of their 25nm 60GB SSDs). (aside, I realize normalized writes don't quite work if OP proportions varies within a product line, sigh)

Out of curiosity, how much NAND do the 80GB and 120GB 320s have?

With the new (to me) knowledge that the 320 40GB has a parity scheme...I don't see the 320 40GB dying for quite awhile.

[Computurd]

http://www.micron.com/products/nand_...clearnand.html

CLEARnand does have the integrated controller on the nand itself, offloading the proc of ECC functions.

[johnw]

But we don't know exactly how much parity data is being written so I won't (not sure it's a safe assumption that the full 8GiB of the sixth die is used for parity).

Right, we do not know. Intel calls it "XOR parity" which is fairly vague (obviously for single parity there is an XOR function, so XOR does not really add any info) and has reportedly described it as "RAID-4 like", but not exactly RAID-4, which makes sense because if it were exactly RAID-4, the parity flash would wear out much sooner than the other flash because RAID-4 parity has to be re-written every time any of the other flash chips are written. So Intel must be using some tricks to avoid wearing out the parity flash too quickly, so we cannot make any assumptions about how much parity data is written. Regardless, I suspect that the block-erases for the parity flash are not included in the SMART attributes.

As for how much flash is on-board 80GB and 120GB Intel 320 SSDs, I am not certain. I don't have any of those models, and I have been unable to find a circuit-board picture for those anywhere on the Internet. The ones I know for certain are 40GB/48GiB, 160GB/176GiB, 300GB/320GiB, 600GB/640GiB. I'd guess the 80GB and 120GB models have either 8GiB or 16GiB extra (one or two 8GiB packages).

[B.A.T]

A 40GB Intel 320 has six 8GiB flash packages on its circuit board. I think that five of them are used normally (five channels of the controller), and the sixth is used for the RAID-4 like XOR parity. So that would be 8GiB extra flash.

My bad. I read an review of the 320 but it might have been the 160GB edition and that one had 16GB for extra sparearea.

[B.A.T]

Morningupdate:

117.8156 TiB
399 hours
Avg speed 89.10 MiB/s.
AD gone from 34 to 32.
P/E 2053.
MD5 OK.

[B.A.T]

As for how much flash is on-board 80GB and 120GB Intel 320 SSDs, I am not certain. I don't have any of those models, and I have been unable to find a circuit-board picture for those anywhere on the Internet. The ones I know for certain are 40GB/48GiB, 160GB/176GiB, 300GB/320GiB, 600GB/640GiB. I'd guess the 80GB and 120GB models have either 8GiB or 16GiB extra (one or two 8GiB packages).

Here is a circuit-board picture of the intel 320 80GB
Link

[johnw]

Here is a circuit-board picture of the intel 320 80GB
Link

EDIT:

I had written a bunch of words about how I could not believe that was a production unit since it only had 80GiB of flash.

Then I looked again. The tenth flash chip, in the lower right corner, is 16GiB

29F16B08CCME1

So the 80GB model has 88GiB of flash (nine 8GiB and one 16GiB)

(insert clip of Christopher Lloyd yelling "88 gibibytes!")



I also see that the 120GB model apparently has six 16GiB and 4 8GiB packages for a total of 128GiB (the link says the gross capacity is 128 MiB, oops!), according to the table on this page:

http://translate.google.com/translat...20_Series_SSDs

[omgFiRE]

Out of curiosity, how much NAND do the 80GB and 120GB 320s have?

Here is a photo of my Intel SSD 320 Series 120Gb (SSDSA2CW120G3): top, bottom (sorry, interface of image hosting site is in russian language). As you can see it has 6x16GiByte + 4x64GiBit = 128GiByte.

[johnw]

Here is a photo of my Intel SSD 320 Series 120Gb (SSDSA2CW120G3): top, bottom (sorry, interface of image hosting site is in russian language). As you can see it has 6x16GiByte + 4x64GiBit = 128GiByte.

Nice! Thanks for posting the pictures.

So, 40GB, 80GB, and 120GB Intel 320 models all have 8GiB of flash for parity. The 160GB model has 16GiB, 300GB has 20GiB, 600GB has 40GiB for parity.

It is also interesting that the only 320 models that have flash on the back of the circuit board are the 160GB and 600GB models.

[Anvil]

164.98TB Host writes
MWI 10
Reallocated sectors, glued at 6.

MD5 OK, 33.06MiB/s on avg for the last 16.5 hours.

edit:

@One_Hertz

Lets hope there are some interesting figures showing up.

[Hopalong X]

johnw and omgFire

Thanks for the info and pictures.
That fills in a lot of info for me on my Intel 320 120GB.

[Vapor]

Very interesting on the 320s' various spare area sizes.

C300 Update

65.644TiB, 78MWI, 1105 raw wear indicator, no reallocated sectors, MD5 OK, 61.85MiB/sec.

[zads]

It is amazing to me that they could dedicate such resources to parity on these devices. I wonder at the sophistication of the parity scheme. For instance, if you look at many raid controllers, etc (same concept essentially as ssd with its controller and nand) running a parity raid set can really cripple write speed in many scenarios. And these full blown raid controllers are with chips that are immeasurably faster than the low wattage chips present on an SSD.

Well the other way you could say it is,

They (Sandforce, Intel, etc) take a type of performance hit (or price hit, depending on how you look at it) to improve error correction capability by more than many many orders of magnitude..
They need to guarantee data integrity better than hard drives for their serviceable lifetime..

[Computurd]

oh yes the gains are definitely worth it, i wish we had more data on the exact parity scheme used, and whether or not it is handled by the controller. amazing they pull it off with such low wattage devices, i wonder what kind of performance could be had if you used an apparently low-overhead parity set, such as these used, on a full powered raid card.

[Ao1]

Here is a photo of my Intel SSD 320 Series 120Gb (SSDSA2CW120G3): top, bottom (sorry, interface of image hosting site is in russian language). As you can see it has 6x16GiByte + 4x64GiBit = 128GiByte.

According to post # 783 an Intel 25nm 64Gb Logical Unit = 69,120Mb or 67.5Gb
64Gb LU = 69,120Mb / 8.4375GiB
128Gb LU = 13,8240Mb / 16.875GiB
6 x 16.875 + 4 x 8.4375 = 101.25 + 33.75 = 135GiB
Of the 8,640 page size 448 byes are for ECC.
Total NAND Capacity = 135GiB
Total NAND excluding ECC = 128GiB
Format Capacity = 111.70GiB?