Intel Optane Memory Plus Low-Cost NVMe QLC Will Rock The Boat

Status
Not open for further replies.

hannibal

Distinguished
The speed difference between optaine and 1”- is bigger than between optaine and MLC. Yep, the later combination would be better, but it would be much more expensive and not so much faster.
 

Giroro

Splendid
QLC has only a 33% increase in capacity, compared to a ~90% decrease in endurance and a ~25% decrease in write speed. It's not worth it.
Should current SLC buffers be replaced with 3d x-point? absolutely.
But Optane isn't going to solve the fact that a consumer SSD with QLC would need so much over-provisioning that you may actually end up losing more capacity than you gain.
 

popatim

Titan
Moderator
Why do you believe they need even more overprovisioning? Aside from allowing space for a fixed SLC buffer that is. Really 100gb buffer is plenty IMO even for a 2tb drive. How often are you doing to write files of that size?
 

pelegbn

Distinguished
Aug 5, 2015
34
2
18,535
Nobody's gonna say something about that gorgeous girl?? I'm sorry, but - Optane what now?! is Optane whatever board is convering the perfection holding it! damn...
 

Giroro

Splendid

A consumer drive would need more over-provisioning because normal desktop users would not be happy if they had to start replacing their drives every 6 months instead of every 5+ years. QLC will not make drives so cheap that they would be disposable in that way.
SLC buffers are there primarily to increase speed, and replacing that buffer with 3D-Xpoint will definitely be a good thing for performance. The buffers do help with endurance, but they don't solve the underlying issue.
Current drives don't wear out because of their SLC buffers, it's because of the MLC/TLC flash that those buffers are dumping into, which is why they need to be over provisioned in the first place. So, I don't think a higher-endurance buffer is going to fix that. If everything else is the same in a drive you replace 3D TLC NAND rated for 3k Program/Erase cycles with QLC rated for 300, then the QLC drive is going to fail in one tenth of the time. I don't see a way for them to solve that issue, except for with more over-provisioning.

The endurance problem is the reason QLC is being marketed to Write-Once Read-Many enterprise workloads and not consumer SSDs.
 

CRamseyer

Distinguished
Jan 25, 2015
425
10
18,795
The Optane drive can absorb all of the random writes if set up to do so. It used to be called Maximized Mode in RST Cache configurations. With all of the random writes going to Optane and then passed to the QLC as sequential data (that barely impact endurance compared to random data over the long term), the drive will last as long as what we have now.

I'm hearing the QLC endurance is actually much better than what most of us (the people commenting here and that actually care about technology) think it has. Part of that is due to stronger LDPC, but some of it is also due to a higher than expected number of PE cycles.
 

CRamseyer

Distinguished
Jan 25, 2015
425
10
18,795
"The endurance problem is the reason QLC is being marketed to Write-Once Read-Many enterprise workloads and not consumer SSDs."

Give me a few hours and I'll toss that right out of the window.
 

Giroro

Splendid


I'm not saying there's no way to design a storage device where the cache doesn't absorb a lot of the writes, like I said, having a cache there helps so we agree on that point. But you have to write to the QLC eventually just like you have to write to TLC. I assume that a drive using this 'maximized mode' with TLC or MLC would also have increased performance, so the performance gap would remain. It does sounds like something that would be interesting when paired with a 10TB spinning disk.

But if it 'used to' be called maximized mode, does that mean that somebody already tried to market the technology and it was rejected? I don't know much about it or why current drives do not operate that way, but would using Optane make it viable for consumer drives when it apparently wasn't when attempted with SLC? Optane is expensive, and the pitch for QLC is that it is supposed to be cheaper, so I don't know if anybody can find a balance where price/performance ever justifies that arrangement.

As for what the P/E of production QLC will be, that is weirdly hard to figure out. I have seen pre-production claims that 3D QLC may be capable of 1000 P/E cycles which would put it near planar TLC in terms of endurance, but still lower in speed - and who knows if they'll get there. Which is still pretty bad considering that planar TLC was already inferior. Granted, 1,000 P/E cycles is a whole lot better than the 100 P/E cycles that they were expecting when planar QLC was being attempted.
I was also having trouble figuring out the exact P/E cycles of current 3D TLC. I've seen a paper from 2015 claiming samsung's "second-gen" 3D TLC would reach 20k P/E cycles compared to 18k from their planar MLC. But I haven't been able to find a good source on if they ever reached that, and what is currently being used in consumer SSDs - I get the impression it lands somewhere in the 3k-5k range. A reduction from 3k to 1k is still very significant, even if it's not as bad as a drop from 20k down to 100.

From my perspective, it just looks like anything they can do to bring QLC to consumer drives with comparable endurance will be too costly, and would also improve TLC/MLC drives so the gap will remain - but it's a good thing too keep trying. That's how we'll get better SSDs overall, at least if they keep the other NAND technologies in production.
Realistically though, if QLC consistently meets the best-case performance claims, then they're just going to rush out a lot of slightly-cheaper low-performing drives with the warranty dropped down from 10 years down to 2 and sell them to people who don't know the difference.
People who are in the market for that kind of thing might be better off picking up a cheap planar TLC drive or two right now while you can still find them on clearance for $30.
 

popatim

Titan
Moderator
SSD's already have dynamic emulated SLC buffers and over-provisioning (where the controller will use any empty space as it see's fit; partitioned or not) so leveraging more as dedicated isn't going to help endurance in the least bit that I can see. Perhaps when the drive nears the too full mark it would if the drive doesn't have the built in ~7 to 15% that the user cannot access to begin with.

I think if we wanted to help endurance we would move the block erase part of writing to empty cells off of the SSD. An SSD can only zero out a block of cells at a time (as opposed to individual cells) and we could save an entire "write" by copying the data out of those used cells to Optane instead of writing it to another part of the SSD. I think we have 5 to 7 yrs before we see this 'synergy' of storage actually occurring in the mainstream.

Also with QLC we are supposed to get much larger drives. So if endurance is rated for 300 cycles, that would still be 1200TB on a 4TB drive. That's nearly 33years writing 100GB every day and not many people write that much. :)
 
Status
Not open for further replies.