SLC & MLC NAND M.2 Consumer SSDs

,

Hello!

As SSD technology progressed, the NAND technology got worse!

I remember the time when 2-bit MLC NAND drives were the norm for high-end consumer SSDs.
Those are gone now, and have been replaced by 3-bit TLC NAND.

I do get that, generally, people want more storage, but there are some who would still prefer high-speed random transfers.

The fact that the best SSDs on the market use the PCIe 5.0 interface, and can hit 14GB/s in sequential transfers, means very little to the average user.

Question:
Why aren’t we seeing consumer high-end SLC & MLC SSDs anymore?

You’ll have a hard time finding them, because it doesn’t appeal to the mass market. That doesn’t mean no manufacturer makes or sells any, but you’ll have to look into niches to find them.

  • All the Intel Optane series, which was a better-then-SLC-NAND mass consumer product before it got canned
  • The PNY LX3030 series, which ran Micron’s 96L QLC NAND in pSLC mode for up to 27 DWPD was short-lived (for the Chia mining fad)
  • The Phison Pascari AI100E series, which claims 100 DWPD, SLC NAND, and low latency will set you back several thousand USD
  • The FREEFLY High-Endurance pSLC SSD for Ember (2.56 TB, pSLC) is advertised as enduring 80 PBW
  • Enterprise (p)SLC drives from Kioxia, Micron, and Solidigm
    • Kioxia’s FL6 series comes the closest to Intel’s Optane performance and endurance and with native 4 KiB NAND page size
  • There’s also all the industrial SLC-based SSDs with very high endurance, but low performance and/or outdated interfaces
  • And surprisingly in the consumer space, a lot of CFexpress cards use pSLC to guarantee steady write bandwidth
3 Likes

As I recall there’s also a new Phison E31T controller and I believe it doesn’t utilize DRAM, and I believe it was shown with TLC NAND.

I see the kind of density and scale enterprise storage seems to be pushing flash towards, and it makes me hopeful for the rest of us normies that we’ll get a little trickle-down soon in the form of denser and more durable SSDs.

The flash in my NAS has got some miles on it, and I’d rather not keep buying SATA SSDs until 2024. :wink:

That PNY Phison Pascari drive is quite interesting.

Nobody is asking for it.

Normal users want a cheap as possible SSD. That is solved by HBM.

Gamers mostly care about sequential read speeds. That is solved by PCI 5.

The high end crowd maybe also cares about sequential async writes. That is solved by DRAM and pseudo SLC cache.

You got it completely wrong :grin:
Sequential read transfer speeds is all that matters to the average user!
That is the only number the average user understands.

High-end SLC and MLC SSDs have faded from the consumer market due to their high manufacturing costs and limited storage capacity compared to TLC and QLC SSDs. Modern NAND technologies prioritize cost efficiency and larger capacities, which appeal to a broader market. Additionally, advancements in TLC and QLC performance, combined with technologies like caching and controllers, now meet consumer needs at lower prices, making SLC and MLC less viable for mainstream use.

Yeah.

I get what you are saying.
So, with other words, the average consumer looks at those big sequential numbers and thinks:
“My new SSD is 10X faster than the old one”, not knowing that, in real world usage, it might not be faster at all!

Average users don’t transfer huge files every day, but they are doing a lot of random tasks.

Even the high-end gen5 SSDs, Crucial T705 & Corsair MP700 Pro, have pathetic random reads in CrystalDiskMark at 4K Q32 T1 (below 100MB), and a lot of latency!

First part yes, second part kind of. Desktop users activity is mostly pretty predictable but, from a drive performance view that mainly looks at things in terms of either 1MB sequential or 4k IOPS, there’s a bunch of reading files of a few hundred kB to a few MB that gets proxied as 1 MB sequential and bunch of small file IO that gets proxied as 4 kB random access.

Neither model’s all that great but, as drive performance tends to be pretty flat above 128-256 kB, the sequential approximation’s often pretty decent. If you’re only going to pick one data point to represent the range from 1 byte to 128 kB, 4 kB isn’t a bad size either. The random part’s more questionable as IO from any one user’s likely pretty correlated. IOPS and random latency are measures that incline more towards cases like databases or badly fragmented files on hard drives than typical SSD activity.

There’s also not much code doing multithreaded IO or single threads capable of more than 1-2 GB/s. So many apps lack the IO performance to saturate a 3.5 or SATA SSD, much less the ~3.5 GB/s limit and lower latency of PCIe 3.0 x4 NVMes, and the ones that can often do bursty IO of a couple GB or less that’s absorbed by OS or file system read prefetch and write caching. Meaning drive latency and bandwidth’s pretty well hidden.

The most common cases for typical(-ish) users to move sizeable amounts of data are probably initial backup syncs, video copies, and transferring image sets. Usually USB, Ethernet, SD card, and hard drive constraints mean those happen below 3.0 x4 performance, though USB 4/Thunderbolt 3 is becoming more prevalent. Since these occasional transfers are usually in the hundreds of MB to tens of GB range, a corollary is the ~400 write and ~600 write minimum lifespans typical of QLC and TLC specs are effectively infinite.

Next to no difference, yup. Speaking as someone who regularly does 1-2 TB sequential reads at the ~7 GB/s limit of PCIe 4.0 x4, it doesn’t mean much to me either. I have to get the 1-2 TB into place first, which is usually network or physical mail constrained (what is the bandwidth of a USB drive in a postal van?) and, in the best case of physically moving 4.0 x4 NVMes between M.2 sockets in different machines, drops to the upper bound of a ~1.5 GB/s cache folding rate once pSLC’s exhausted.

Once that’s done there’s a few initial processing steps that do 7 GB/s with subsequent ones running out of desktop cores at 5-6 GB/s, getting stuck at ~3 GB/s because that’s where the code libraries they need to call through peak with 24-32 threads, or hitting other constraints. With a 5.0 x4 it looks like the initial steps would run out of cores at 10-12 GB/s. But, if it’s not an SM2508 or E31T drive, probably I’d have to throttle the workload below that to avoid overheating even with good airflow and the drives under upper spec NGFF heatsinks.

I think 14 GB/s is not out of the question once 16 core Zen 6 is available. But reading 2 TB at 12 GB/s takes 167 s. At 14 GB/s it’s 143 s. Either way I’ll have switched to doing some other task while the job’s running. It’s unlikely I’ll happen to check back somewhere between 143 and 167 s.

Diminishing returns are diminishing. The marketing numbers, not so much. :smiley:

100 MB/s is 25,000 4k IOPS. Not many real world workloads where one thread generates that many random requests for long enough that it matters. Like, when was the last time you said hey, I have 100 GB on this drive, I gotta have a 0.1% random sample of that data, and ZOMG I absolutely positively need this app code to have finished doing that five seconds ago?

Also, I think there’s an erroneous conflation of Q1T1 and Q32T1 here. At Q1 DiskSpd/CrystalDiskMark does sequential IO, meaning the thread blocks until the 4 kB comes back, then issues the next IO. At Q2 and up it’s async, meaning the thread issues multiple requests and then picks back up when the first of them returns. In IOPS bottlenecked workloads that are implemented for performance (e.g. databases) the code’s usually multithreaded async, so Q32T16’s likely a more relevant measure.

1 Like

I think the gamer sees the PS5 with its high sequential read speeds for games and says to himself “Maybe I don’t need direct storage today, because I am not playing ratched and crank on PC, but I wanna be future proof”

Average users would be happy with a 1TB QLC drive as OS drive, as long as it is cheap.
That is more then enough to open Excel, Outlook and Teams.

I agree with what you wrote, but databases and async would be a desaster :wink:

1 Like

My whole point is that, first time when I’ve replaced my old Seagate HDD with a Kingston MLC SATA SSD, the difference was very notable!

Then, later on, when I’ve replaced my MLC SATA SSDs with MLC M.2 NVMe PCIe 3.0 x4 SSDs, the difference wasn’t very notable.

So, I don’t expect going from MLC PCIe 3.0 to TLC PCIe 5.0 will make a huge positive difference, now that the NAND itself gets more bits per cell.

3DX-Point (Optane) should’ve been the next step in drives evolution, but Intel & Micron stopped the development of those!

QLC is terrible, and PLC will be worse than the old HDDs.

But, I agree with you:
Most people will say it’s good technology.
(Just as I was hearing some dumb guys saying, in the 2000s, that MP3s sound better than audio CDs) :rofl:

512 GB’s probably plenty, maybe 256. The personal laptop I’m typing this post on has a 512 with 240 used and I could easily free enough of that for write amplification not to be much of an issue. At work we use 2 TB system drives because ~0.7-1.1 TB for OS+programs+small user data’s common. One system drive is still a SATA SSD and others are 3.0 x2, 3.0 x4, or 4.0 x4 depending on the build gen. Nobody really cares but the 4.0 x4s are nice for occasional small data tasks like reprocessing ~50 GB, though that’s limited to workflows where I’ve written tools which can utilize 4.0 x4s.

Aside from tending to be a bit slower a limitation I’ve noticed with HMB drives is IO to them on Windows 10 22H2 (and presumably 11, which we’re avoiding as much as practical) isn’t as memory efficient as DRAMed drives. This tips some workloads into being dual channel DDR4 limited rather than drive limited.

Async’s fine. Usually when this comes up async IO threading’s being erroneously conflated with the zfs tendency to call write back IO async. Sometimes the async thread handoff to the kernel needs to explicitly specify write through, sometimes it’s the default. Trivial either way.

Yeah, diminishing returns are down to negligible in typical workloads for the code and application demand reasons mentioned. Bits per cell’s not an immediate factor there, though, and to the best of my knowledge there’s insufficient data on drive aging to tell if QLC’s more prone to slowdowns than TLC.

Optane’s not cost competitive with pSLC and, roughly speaking, a one gen newer standard NVMe benches about the same. I don’t follow Optane closely but, from pricing I’ve seen mentioned around Level1 threads lately, it looks like even used PCIe 3.0 Optane is higher cost per TB than the pSLC on a new, upper end PCIe 4.0 NGFF. Also hard for 3DX to be idle power competitive.

With the pricing I can get, QLC HMB drives tend offer around a third higher price-performance than TLC DRAMed drives at the expense of a similar reduction in operating lifespan. Since the performance and lifespan hardly matter in most use cases it makes sense to choose the lower cost drives.

Much to the disappointment of storage enthusiasts everywhere. :smiley:

That recently changed in my opinion.
16GB for hybernate, 16GB for OS, 16GB for OS upgrade, 20GB for Outlook cached mails, programs have become huge and people hoard lots of pictures.
It is not that 512 is impossible, people are just too lazy to offload files in Dropbox, OneDrive, Nextcloud and GDrive.

Winodows 24H2 got rid of the 64MB HBM limit.

It is in some cases, but never for databases. Databases should be ACID.
But it isn’t up to anyway, since you do not decide if something is async or not (unless for networks shares or if you like to live dangerously with ZFS)

QLC is just a technology. There are 120TB Solidigm SSDs that offer decent performance and TBW.

This is not a correct description of coding IO calls, whether database or otherwise. Async threading’s fine for ACID.

1 Like

Yeah.
I am a tech enthusiast.
So, I do take into consideration the option to buy a used Intel Optane drive.

The average lay person likes big numbers, and tends to make decisions based on two or three figures (price v. capacity v. speed). Enthusiasts/prosumers know what is garbage (3D XPoint v. SLC v. pSLC v. MLC v. TLC v. QLC v. PLC). But the experts calculate the thresholds where certain features matter and at what risk/cost over the long term. :wink:

We all have some kind of products which we do not understand very well and buy based off of what the marketers would like us to know. That’s why subpar products exist. :slightly_smiling_face:

They did make an Optane SSD with idle power in the milliwatts range, but that died with the 800P series’ discontinuation. All subsequent Optane SSDs were for enterprise/data center use.

1 Like

It’s not always that people are lazy.
It’s also the fact that some users prefer having their personal stuff at hand, and not on a public server.
The cloud is someone else’s computer!

There are people who bought media content, or software, on certain platforms (stuff that couldn’t be downloaded).
Later on, the platform removed the stuff that they have payed for, without offering any compensation, or the option to download the stuff that they’ve payed for!
(Netflix and Sony are such examples).

True

Indeed. Though I often wonder how much of that’s the result of marketing practices training folks to think in such terms.

I’ve thought about Optane but can’t find a worthwhile use. Run some refurb Exos 2X18, though, as it’s US$ ~12/TB.

I have an old Corsair Force 3 MLC 120G SATA SSD and a newer 240G TLC PNY SATA SSD among others. After 2 years of use the newer SSD shows ~3% of lifespan gone. The Corsair after 12+ years of use shows 2%.

I also have some HDDs for data in my home server and do regular backups of around 3-4TB of files to an external HDD. In my main machine I have two 1TB NVMes, one Kingston Fury Renegade and a Goodram PX500 Gen 2 and used SATA SSD as well. Almost all drop to about 80MiB/s after the initial burst (50-100GiB of data), while the Renegade Fury can sustain writes in ~8x of that and HDDs sometimes in 150+MiB/s.

While for gaming a good NVMe is the best, for other of my use cases it soesn’t make much of a difference.