3D QLC Woes - Manufacturers Fighting to Get Yields Above 50%

[Raevenlord]

3D QLC (quad-level cell) is the latest, manufacture-ready technology to grace the NAND panorama, with promises of increased density over 3D TLC (triple-level cell), thus bringing pricing per GB even lower. However, as with all wafer-based PC components, yields are an extremely important part of that process. Cost reduction can only be attained if manufacturing allows for a given percentage of a wafer to be fully functional and without defects that compromise its feature-set or performance. However, as cell design becomes more complex in a bid to increase areal density, yields have taken longer to mature.

According to DigiTimes, 3D TLC yields have only gotten off the ground in the beginning of this year - right around the time companies were rolling out their 3D QLC designs. And if TLC took longer than expected to achieve respectable yields, it seems that QLC memory will take even longer - we already knew that the Intel-Micron venture on QLC was facing less than 50% yields, but DigiTimes has now extended this struggle to what seems to be the entire NAND manufacturing industry (Samsung Electronics, SK Hynix, Toshiba/ Western Digital and Micron Technology/Intel). The result? Expected price fluctuations in the beginning of 2019, as predicted production volume fails to meet both projected and actual demand, with 3D TLC supplies having to cope with increased market demands.



View at TechPowerUp Main Site

 

[HTC]

Translation: price hikes ...

 

[TheLostSwede]

That might be a good thing. I'm in no rush getting anything with QLC in it, not even a memory card. This stuff really needs to prove it's reliable.
I skipped TLC for SSDs and only just now got my first 3D TLC SSD. QLC just looks too iffy to me.

Translation: price hikes ...

In other words, if you need an SSD, get one now, while they're cheap-ish.

 

[trparky]

QLC just looks too iffy to me.

Me too.

 

[Hood]

QLC NAND may be necessary to get prices lower, but could have the undesired effect of lowering perceived reliability of SSDs. In the past, the questionable reliability of NAND and controller chips has caused adoption rates to suffer, along with high prices, and short usable life may cause the same problem when QLC is more common.

 

[Vayra86]

QLC NAND may be necessary to get prices lower, but could have the undesired effect of lowering perceived reliability of SSDs. In the past, the questionable reliability of NAND and controller chips has caused adoption rates to suffer, along with high prices, and short usable life may cause the same problem when QLC is more common.

While true, in terms of write endurance all drives and all NAND types have proven to (far) exceed the specs.

I think the main problem with QLC is performance. The Intel 660p for example, its a weird product. Its a slow as the average SATA SSD and is considered competitive because its now in M2 and has the potential for higher sequential speeds and IOPS. In other words, there are barely any changes when compared to a regular SATA SSD which is as cheap, a bit slower in IOPS but much more flexible in terms of placement and much faster at sustained write operations. And here comes the kicker... we're talking about a DRAM supported M2 SSD as well. Not exactly cost effective if you place it next to no-DRAM SATA SSDs...

The bottom line, if QLC doesn't significantly drop in price (as in, below $0.10/GB) the changes are so minimal for most people its not even worth taking the lower endurance rating for granted. There's just nothing substantial to justify the small price gap compared to a regular SATA SSD. The only justification right now is when you place it next to M2 SSDs exclusively... but who really does that? Its all storage, with a certain amount of performance. PCIe doesn't magically make things better and it does eat valuable lanes.

 

[hojnikb]

>However, as cell design becomes more complex in a bid to increase areal density, yields have taken longer to mature.
QLC on it's own does not increase complexity of a cell design. Fundamentally, a cell is exactly the same for SLC, MLC, TLC or QLC type NAND.

The supporting circuits to enable increased error correction due more fine grained voltage differentiation (in case of QLC there are 16 different voltage states, which means a very narrow voltage range for each cell value). do have to be beefed up for somewhat reliable operation (obviously controller support also plays a vital role here).

 

[efikkan]

I'm less concerned with production yields, the real problem with SSDs is long-term reliability. TLC is bad enough already. We desperately need a more reliable technology.

Makers can say all they want about TB written and spare cells, once these start to become unreliable it means data loss and an unstable device.

 

[DeathtoGnomes]

Love a good fantasy manufacturer drama.

 

[CheapMeat]

I'm less concerned with production yields, the real problem with SSDs is long-term reliability. TLC is bad enough already. We desperately need a more reliable technology.

Makers can say all they want about TB written and spare cells, once these start to become unreliable it means data loss and an unstable device.

Well, you could go Optane then if your main concern is endurance. The endurance beats out even SLC NAND. Plus it doesn't slow down the fuller it gets.


From sites that have done torture tests, personally, I'm not worried at all about 3D TLC NAND. I'd only worry about older planar NAND. But even then you can compensate by just getting a bigger drive.

 

[StrayKAT]

Translation: price hikes ...

Hopefully the Intel model (660p) won't get too expensive. It's shipping later this month, but Newegg has a 1tb M.2 model for preorder at $189. That's not bad. edit: Didn't they come up with QLC themselves? Maybe that's why it's first in production and isn't price hiked yet.

https://www.newegg.com/Product/Product.aspx?Item=N82E16820167462

 

[fynxer]

NO WAY all manufacturers have the exact same problem, this is bull sh*t on the highest level for real.

 

[TheGuruStud]

I'm less concerned with production yields, the real problem with SSDs is long-term reliability. TLC is bad enough already. We desperately need a more reliable technology.

Makers can say all they want about TB written and spare cells, once these start to become unreliable it means data loss and an unstable device.

Baloney. And the 840 is an old drive. They'll never fail with normal use and will far exceed the life of multiple PCs. If you're going to use them heavily, then obviously you buy MLC. https://techreport.com/review/27909/the-ssd-endurance-experiment-theyre-all-dead

NO WAY all manufacturers have the exact same problem, this is bull sh*t on the highest level for real.

They all have the same problem of collusion.

 

[StrayKAT]

Optane and enterprise drives last much longer than usual... but you'll pay for it.

 

[efikkan]

Baloney. And the 840 is an old drive. They'll never fail with normal use and will far exceed the life of multiple PCs. https://techreport.com/review/27909/the-ssd-endurance-experiment-theyre-all-dead

Just lately, my 1.5 year old Intel 540s 480GB became unusable. The SMART data doesn't show total written data, but I'm pretty sure it's less than 1 TB written.

I've seen so many SSDs and HDDs fail, I can tell you it happens a lot.

 

[StrayKAT]

Just lately, my 1.5 year old Intel 540s 480GB became unusable. The SMART data doesn't show total written data, but I'm pretty sure it's less than 1 TB written.

I've seen so many SSDs and HDDs fail, I can tell you it happens a lot.

Good to know.. I was thinking of getting one of the 545s's.. it's probably not much better. I just want a disk for playing around with Linux (which mean a lot of writes, if I experiment with distros or build from source). Wouldn't want it to crap out too quickly.

 

[Deathy]

Just lately, my 1.5 year old Intel 540s 480GB became unusable. The SMART data doesn't show total written data, but I'm pretty sure it's less than 1 TB written.

What is "unusable". How can you see SMART data if it is "unusable". SDDs fail. But mostly, not because of cell write endurance running out. If the SSD reaches all the allocated pe cycles, it goes into a controlled state where it allows the data to be retrieved for some time but not allow new data to be written.

I've seen so many SSDs and HDDs fail, I can tell you it happens a lot.

And my last failed drive was a 20GB HDD in 2005, your point is what? I even recovered a 2TB 2.5" HDD that got a bit of moisture due to a water cooling accident. My 60GB OCZ Agility still works, my 120GB Vertex 2 still works, my 500GB 840 still works, my brothers 2 830s in RAID 0 still work....

I'm excited for QLC, not for my desktop needs, but I wouldn't mind a file server based on QLC SSDs for a WORM like usage.

 

[StrayKAT]

I'll take a wild shot and say you should Trim if you haven't.. that makes some data usable again after neglect. But I wouldn't be surprised if a cheapo drive craps out on you.

 

[efikkan]

Good to know.. I was thinking of getting one of the 545s's.. it's probably not much better. I just want a disk for playing around with Linux (which mean a lot of writes, if I experiment with distros or build from source). Wouldn't want it to crap out too quickly.

Hard to tell, really.
Almost all of my SSDs are Intel, but I've set up machines with Samsung etc., but they are not better. Samsung 950/960 are known to have firmware issues on Linux. But I've never seen a recent Intel (520/530/535/540) having firmware issues in Linux.

For the record, my bad Intel 540s was running Ubuntu 16.04, not that I have any evidence supporting that the OS killed it. It's getting random bad sectors all over, which I assume is bad flash.

What is "unusable". How can you see SMART data if it is "unusable". SDDs fail. But mostly, not because of cell write endurance running out. If the SSD reaches all the allocated pe cycles, it goes into a controlled state where it allows the data to be retrieved for some time but not allow new data to be written.

Mostly, I'm talking about rapidly growing SMART error counts.
For instance, my bad 540s saw random files and folders all over become unreadable. It's still not completely dead, but not usable.

 

[TheGuruStud]

Just lately, my 1.5 year old Intel 540s 480GB became unusable. The SMART data doesn't show total written data, but I'm pretty sure it's less than 1 TB written.

I've seen so many SSDs and HDDs fail, I can tell you it happens a lot.

Don't buy intel junk, problem solved lol. The point is that they're many times more reliable than a mechanical. Unless there's a design flaw (like the old shitty controllers and fw bugs) or it's straight defective, the odds are it will last many years, and likely the lifetime of that usable capacity. It seems all of the issues have been ironed out and only the no-name chinesiums have problems.

I haven't seen any SSD issues and only heavily used mechanicals have failed in general...unless we're adding in crapgates.

 

[StrayKAT]

Don't buy intel junk, problem solved lol. The point is that they're many times more reliable than a mechanical. Unless there's a design flaw (like the old shitty controllers and fw bugs) or it's straight defective, the odds are it will last many years, and likely the lifetime of that usable capacity. It seems all of the issues have been ironed out and only the no-name chinesiums have problems.

I haven't seen any SSD issues and only heavily used mechanicals have failed in general...unless we're adding in crapgates.

It varies with Intel. They seem to have everything from the best drives (900/905p) to some of the worst. And like I said, if I'm not mistaken, they (and Micron) are the first to get QLC on the market...

 

[hat]

So suddenly either yields are much worse than they were previously for "some reason", or yields have always been this way, and it's just now suddenly a problem.

 

[lexluthermiester]

QLC just looks too iffy to me.

Same here. Reliability is highly questionable. I'm inclined to pass on anything with QLC in it until they can prove the technology is robust enough for practical usage scenario's.

 

[R0H1T]

Well, you could go Optane then if your main concern is endurance. The endurance beats out even SLC NAND. Plus it doesn't slow down the fuller it gets.


From sites that have done torture tests, personally, I'm not worried at all about 3D TLC NAND. I'd only worry about older planar NAND. But even then you can compensate by just getting a bigger drive.

According to some specs or projections, when was the last time you saw an endurance test with SLC NAND like ~ SSD Endurance Test: BX200 dead after 187TB, 850 Pro after 9.1PB

In essence good NAND, SLC or MLC, will outlast most human attempts to exhaust their PE cycles.

So suddenly either yields are much worse than they were previously for "some reason", or yields have always been this way, and it's just now suddenly a problem.

The QLC yields were never disclosed, probably because there's only a handful of drives using it.

 

[trparky]

So if QLC is turning out to be such a pain in the... rear end, why keep trying for it? Expand on TLC NAND, keep adding more layers, etc.