Question MLC SSDs are on their way out?

[mikeymikec]

After a thread about the 980 PRO, I thought I'd start hunting around for what I'd consider changing to after the 860 PRO, however from the reading I've done so far plus this review:

Best SSDs: May 2021

www.anandtech.com

MLC seems to be a bit of a rarity?

 

[razel]

860 PRO is a SATA SSD. I'm be happy moving onto nearly any budget NVME SSD such as Intel 660p, 665p, WD 550 or Kingston A2000 that are in the under the $.10 a GB range. I wouldn't bother with the 980 Pro. Just not worth the price or real world performance benchmarks. For the price of one 1TB 980 Pro you can get 2TB of the budget drives. Far better value.

 

[thecoolnessrune]

I'd say so, even the vast majority of write intensive Datacenter SSDs being released right now are using TLC. The endurance, and performance for has been there for a while for Enterprise, so most of it is certainly good enough for consumers in most cases.

I was looking at my Samsung 960 EVO the other day, and in the 3 years I've had it, I've only written a bit over 15 TB to it. Certainly for most consumer workloads, TLC has more than enough endurance to satisfy, especially in the higher capacities.

 

[mikeymikec]

One of my worries with MLC disappearing is that it's part of a move to encourage repeat purchases. Who wants to sell a drive that a user might be happy for 10-20 years with? IMO it's the reason why computers with hard drives are still being sold, because they perform a lot crappier and users will want to upgrade.

We could do with another set of endurance trials from say techreport for TLC drives this time.

 

[UsandThem]

Some TLC NVMe drives this past generation came with higher TBW warranties than the 970 PRO offered (example BPX PRO).

I admit I was initially concerned when TLC NAND first began appearing on SSDs, but now that the process has since matured over many years, I don't have any worries about it. I imagine most people will replace their SSDs because they are "obsolete" and not because they failed.

 

[Insert_Nickname]

Who wants to sell a drive that a user might be happy for 10-20 years with?

I have a 4 year old Intel 600p in use. It's currently at 7.64TB written. It's not like I've spared the drive in any way. The 512GB 600p is rated for 288TBW, so I should be set for the next 150-odd years. Somehow I doubt write endurance will be a problem in this case...

(sorry about the tongue-in-cheek, but I couldn't resist... )

 

[mikeymikec]

Yep, I calculated my 840 PRO's life expectancy to be a similarly absurdly high figure as well, but more bits per cell means more work per cell, which means that any manufacturing flaws will be put under greater stress which means increased likelihood of failure before MTBF.

I've seen one PRO drive fail in warranty (light load) and one other SSD, and second-hand experience of another one.

Techreport's endurance tests inspired faith in me in SSDs after the era of having to be careful about wear-levelling, actual independent testing showed that some drives would last a very long course. It might be that with the newest manufacturing techniques used in more modern drives that they last even longer than the best MLC drives from the previous round of tests, who knows. However it might be that they don't fare as well, perhaps because of the nature of TLC or maybe some other manufacturing cheap-out.

 

[Insert_Nickname]

I've seen one PRO drive fail in warranty (light load) and one other SSD, and second-hand experience of another one.

It happens. I've seen brand-new enterprise HDDs fail within the week, and the cheapest c*** drive doing 10 years almost 24/7. There is just no way to predict failure.

To be fair, SSD are still a heck of a lot more reliable then mechanical HDDs. It's anecdotal of course, but I think I've had two SSDs fail on me in 10 years. HDDs have been 20+.

Techreport's endurance tests inspired faith in me in SSDs after the era of having to be careful about wear-levelling, actual independent testing showed that some drives would last a very long course. It might be that with the newest manufacturing techniques used in more modern drives that they last even longer than the best MLC drives from the previous round of tests, who knows. However it might be that they don't fare as well, perhaps because of the nature of TLC or maybe some other manufacturing cheap-out.

3D NAND -should- be more robust compared with later generation planar NAND, but nobody really knows yet. 3D NAND just hasn't been around that long yet.

 

[UncleJoe1985]

I suspect the actual write cycles are much better than what is advertised due to the huge SLC caches. The 980 PRO has ~114 GiB of it. If your writes have good locality (e.g. you write to the same location over and over), then only the SLC will get used. Consider a 980 PRO, which has an endurance of 600 drive writes. I assume that's calculated based on the worst case of filling the whole disk, causing every cell to get overwritten each pass. But if you don't fill your whole drive and the wear leveling was perfect, then you will be able to use every cell to their maximum endurance:

11.4% * 50k cycles + 88.6% * 600 cycles = 6232 cycles

Can any one confirm this?

 

[Billy Tallis]

I suspect the actual write cycles are much better than what is advertised due to the huge SLC caches. The 980 PRO has ~114 GiB of it. If your writes have good locality (e.g. you write to the same location over and over), then only the SLC will get used. Consider a 980 PRO, which has an endurance of 600 drive writes. I assume that's calculated based on the worst case of filling the whole disk, causing every cell to get overwritten each pass. But if you don't fill your whole drive and the wear leveling was perfect, then you will be able to use every cell to their maximum endurance:

11.4% * 50k cycles + 88.6% * 600 cycles = 6232 cycles

Can any one confirm this?

The 980 PRO does not have 114GiB of SLC. It has 1TB of TLC. It will use varying amounts of that as SLC to provide a cache of between 6GB and 114GB. How much TLC gets treated as SLC, and which physical blocks of NAND get treated as SLC will vary over time and depending on usage patterns. Like most consumer TLC SSDs, data that is initially written to SLC will generally be compacted and written out to TLC blocks during idle time, so mere spatial locality of writes is not enough to ensure that your data only gets written to SLC and never moved to TLC.

Since the specific physical flash blocks that get used as SLC changes over time, there is no "SLC section" that can be regarded as having many thousands of P/E cycles. A P/E cycle on a SLC block contributes to the wear of the general pool of available NAND flash, and the drive will cease functioning as soon as the flash is too worn to be used as TLC.

Most consumer QLC drives are much less aggressive about flushing the SLC cache during idle time, so choosing to only use the first quarter of a QLC drive will get you a lot of SLC-like characteristics for performance and wear.

The Enmotus FuzeDrive SSD is a QLC drive with an actual static SLC partition the way you imagine it, with blocks that are permanently treated as SLC for the lifetime of the drive and mapped to a specific range of LBAs, and the SLC portion and QLC portions of the drive have independent wear leveling processes.

 

[UncleJoe1985]

The 980 PRO does not have 114GIB of SLC

Heh, I did think that was too good to be true. Thanks for clearing that up.

So that means when using TLC as SLC, you're spreading the writes over 3 times the # cells. That's a write amplification factor of 3x ! Terrible. Then when you flush to TLC, that's 4x.
I saw a quote of 3D TLC NAND having 1.5 to 3k write cycles. So that explains how it drops to 600 cycles.

There should be a way to turn off or minimize SLC caching. Reducing drive life by 3x just to increase write speeds from 2 GiB/s to 5 GiB/s doesn't seem worth it.

 

[thecoolnessrune]

Heh, I did think that was too good to be true. Thanks for clearing that up.

So that means when using TLC as SLC, you're spreading the writes over 3 times the # cells. That's a write amplification factor of 3x ! Terrible. Then when you flush to TLC, that's 4x.
I saw a quote of 3D TLC NAND having 1.5 to 3k write cycles. So that explains how it drops to 600 cycles.

There should be a way to turn off or minimize SLC caching. Reducing drive life by 3x just to increase write speeds from 2 GiB/s to 5 GiB/s doesn't seem worth it.

No worries! it doesn't actually have that sort of effect because of how it's done. In fact, it's the opposite, SLC mode can actually extend the lifespan of a NAND cell, by as much as 20x according to vendors like Hyperstone. This is because of the preciseness of Voltage states needed. Because SLC only needs to maintain 2 states (a 0, or 1 per cell), NAND can be very "roughly" programmed. They don't need super exact state voltages, because there's a large amount of buffer. On the opposite end is TLC, which needs to hold 8 voltage states, and QLC which needs to hold 16 Voltage States. As NAND is programmed and wears, it requires higher and higher voltages to program. While SLC has a huge amount of room to program between "low and high" voltage states, TLC and especially QLC have a tiny amount of buffer room in comparison. Indeed, on advanced controllers, NAND that has exceeded its P/E Cycles to be used as TLC or QLC can still live perfectly functional lives as SLC NAND. Anandtech actually wrote a great synopsis on this process back in 2012, when there were only just rumblings of TLC NAND: https://www.anandtech.com/show/5067/understanding-tlc-nand/2

While there is write amplification involved with going from SLC to TLC, not all write amplification is created equal. NAND as SLC is useful for a lot longer than it is at TLC. NAND written as SLC may be good for over 100,000 program / erase cycles, whereas in TLC mode it may only be good for 5,000 P/E Cycles

 

[UncleJoe1985]

I did read about pseudo SLC. Sure it lasts much longer and is faster to read/write, but you're losing 3x the space. For most people, once cells can no longer hold 3 bits, then the drive isn't usable. Unless manufacturers can convince people that it's natural for the disk to shrink in size as it ages.

 

[mikeymikec]

Reducing drive life by 3x just to increase write speeds from 2 GiB/s to 5 GiB/s doesn't seem worth it.

IMO there will be some niche that will fancy the candle that burns twice as bright for half as long.