More Details on SK Hynix 321-Layer NAND Flash Appears at the Flash Memory Summit

[TheLostSwede]

Courtesy of an SK Hynix keynote speech at the Flash Memory Summit, we now have a few more details about its upcoming 321-layer NAND Flash. PC Watch Japan who attended the industry event shared some pictures from the keynote which adds some crucial details that were missing from last week's press release. SK Hynix's officially shared performance figures tell us that we should expect up to 12 percent faster program performance, which should be the write performance and up to 13 percent improved read latency. Both of these performance metrics will obviously depend on the SSD controller the NAND is paired with, the related firmware on said controller and so forth.

PC Watch Japan also quotes a program throughput of 194 MB/s, which is 26 MB/s improvement over SK Hynix 176-layer NAND and currently the highest known program throughput of any announced NAND Flash. That said, Kioxia is expecting to hit 205 MB/s with its next generation of 300 layer NAND. SK Hynix also claims 10 percent better read power efficiency, which is really neither here nor there when it comes to modern SSDs, unless we're talking server level SSDs with a dozen of these NAND chips or more. Rather than going with two stacks of 150 plus layers each, SK Hynix went with three times 107 layer stacks, which should be compared to their current 238 layer product which has two stacks of 119 layers. This made the new NAND package easier to produce and should in the long term result in higher yields. Each NAND package is expected to deliver a memory density of 20 Gbit per square millimetre or more, which is almost twice that of its 176-layer NAND.



View at TechPowerUp Main Site | Source

 

[bonehead123]

Soo..... to summarize:

gather up all your spare arms, legs, kidneys, 1st born childs AND your banker's gold & platinum cards, cause this stuff aint gonna be cheap

 

[mechtech]

I’d be more interested in PE cycles and endurance. (TBW)

 

[TheLostSwede]

I’d be more interested in PE cycles and endurance. (TBW)

Considering there are more layers, it should be up, but I have no idea by how much/little. The article doesn't mention it and considering availability isn't expected until early 2025, it might just be that the first samples they've made aren't good enough to test things like this. Besides, it also takes time to do stress testing, so we might not find out until next year.

 

[Denver]

All very nice until you see that it will only start to be produced in 2025. Smells as bad to me as intel's claims "By 2025 we will overtake TSMC, and AMD will be so far behind that we'll see it in the rear view"

 

[AnarchoPrimitiv]

I want someone to make a 256GB ALL SLC SSD that I can use for an OS drive. Dapustor makes all SLC drives for enterprise, but it'd be cool if someone would make small enthusiast SLC drives.

All very nice until you see that it will only start to be produced in 2025. Smells as bad to me as intel's claims "By 2025 we will overtake TSMC, and AMD will be so far behind that we'll see it in the rear view"

Yeah, Intel is already spending more on R&D than TSMC and AMD combined, so I can't imagine how much more cash they'd have to burn through to overtake them.

 

[Pizdarenkowitch]

Why someone should care about this if ULTRA-RAM is behind the coroner ???

P.S I'm talking about this : https://quinas.tech/wp-content/uploads/ultraram-quinas-brochure.pdf

 

[mashie]

Is this dense enough to enable 16TB M.2 drives?

 

[TheLostSwede]

Is this dense enough to enable 16TB M.2 drives?

Most likely not.

 

[Wirko]

Most likely not.

Most likely yes. Even more likely, there's a huge BUT attached to it.

8 TB M.2 SSDs exist, each of them packs 128 x 512 Gbit TLC dies or 64 x 1 Tbit QLC dies in eight packages. But if you also care about price, you can buy a 16 TB drive in a more comfortable U.2 or U.3 shape for not much more money than an 8 TB M.2 drive.

By analogy, I expect to see incredibly expensive 16 TB M.2 SSDs with 128 of these new memory chips inside, while physically larger variants will be far cheaper (and cooler too).

 

[Tomorrow]

Most likely yes. Even more likely, there's a huge BUT attached to it.

8 TB M.2 SSDs exist, each of them packs 128 x 512 Gbit TLC dies or 64 x 1 Tbit QLC dies in eight packages. But if you also care about price, you can buy a 16 TB drive in a more comfortable U.2 or U.3 shape for not much more money than an 8 TB M.2 drive.

By analogy, I expect to see incredibly expensive 16 TB M.2 SSDs with 128 of these new memory chips inside, while physically larger variants will be far cheaper (and cooler too).

4TB models already use 1TB TLC chips. The first one i bought has two 1TB chips and presumably two 1GB DDR4 cache chips on each side.
The main limiting factor are the controllers themselves being limited to 8 channels and 8TB for mainstream SSD's.

4TB seems to be finally affordable but 8TB is still too expensive and there are no mainstream 16TB M.2 options nor do i expect any to appear soon.

 

[Wirko]

4TB models already use 1TB TLC chips. The first one i bought has two 1TB chips and presumably two 1GB DDR4 cache chips on each side.
The main limiting factor are the controllers themselves being limited to 8 channels and 8TB for mainstream SSD's.

4TB seems to be finally affordable but 8TB is still too expensive and there are no mainstream 16TB M.2 options nor do i expect any to appear soon.

I tend to avoid the word "chip" because it can mean "die" to one person and "package" to another. Many silicon dies are stacked inside each black epoxy package, I checked the TPU database for existing 8 TB M.2 drives and found the two arrangements I mentioned.
Controllers are a limiting factor too of course, but sometimes it's hard to find reliable data - for the SM 2508 (PCIe 5.0), the limit seems to be 8 TB.

 

[DutchTraveller]

I recently bought a 2TB SN570 but the production-date is almost a year ago. It has only one Nand package (of 2TB) and the controller on it.
Apparently it is cheaper to use 16 1Tbit dies in one package than 32 512Gbit in 2 or more packages.
There is space for 2 packages on top (giving 4TB) and 2 on the bottom (which in total results in 8TB).
But according to data found on the internet the controller is already maxxed out at 2TB.

So another controller is needed to produce 4 or 8TB drives but the current Nand (1Tbit dies) is not the limiting factor.
Should even be cheaper per GB than the smaller capacities but as always the large capacities carry a premium.
Recently Lexar introduced a fast and cheap 4TB M2 drive.
With the current glut of Nand maybe more large drives will be introduced at interesting prices.

You can always go for an enterprise drive like the PM9A3 which is available with a U2 interface in capacities up to 15.36TB for reasonable prices.
There are interface cards to use them in a PCIe slot.

 

[TheLostSwede]

I want someone to make a 256GB ALL SLC SSD that I can use for an OS drive. Dapustor makes all SLC drives for enterprise, but it'd be cool if someone would make small enthusiast SLC drives.


Yeah, Intel is already spending more on R&D than TSMC and AMD combined, so I can't imagine how much more cash they'd have to burn through to overtake them.

pSLC you mean?
Like this? US$82 for 320 GB.

达墨 pSLC 旗舰 SSD 公布：1280GB 42000 TBW，售价 2299 元 - IT之家

达墨日前预热的消费级 pSLC 旗舰 SSD 今日公布，型号为 Virgo U，1280GB 2299 元，预计将在不久后公布。

www.ithome.com

 

[Selaya]

I want someone to make a 256GB ALL SLC SSD that I can use for an OS drive. Dapustor makes all SLC drives for enterprise, but it'd be cool if someone would make small enthusiast SLC drives.
[ ... ]

sounds like optane's what you're looking for tbh

 

[TumbleGeorge]

The main limiting factor are the controllers themselves being limited to 8 channels and 8TB for mainstream SSD's.

Yes most of today modern mainstream controllers are limited to 8TB, but E-26 from Phison support up to:

Up to 8 Channels with 32 Chips Enable (CE)
– Capacity up to 32TB

From storage reviews.
This has been public knowledge ever since it was introduced.

 

[Tomorrow]

Yes most of today modern mainstream controllers are limited to 8TB, but E-26 from Phison support up to:

From storage reviews.
This has been public knowledge ever since it was introduced.

And currently only Crucial T700 has 4TB option. Others are 1TB and 2TB. No other 4TB and no 8TB options. Also based on T700's price the bigger capacities based on this controller would cost roughly as much:
8TB: ~1300€
16TB: ~2600€
32TB: ~5200€

Personally i doubt we will ever see above 8TB utilized for Gen5 even if the controller itself supports it. The market for 16TB 2500+ price M.2 SSD's is very small.
Especially considering that first gen Gen5 models dont offer that much of a performance increase (perf per W may actually decrease due to higher power consumption on certain workloads) and that anyone can get 4TB Gen4 models for under 250€ now.

Manufacturers should focus on getting 8TB Gen4 to the masses. Currently it's still overpriced and should start from 500€ but starts from 800€ (M.2, yes U.2 2,5" does start from 500€ but U.2 compatibility on desktops is nonexistent).

 

[DutchTraveller]

And currently only Crucial T700 has 4TB option. Others are 1TB and 2TB. No other 4TB and no 8TB options. Also based on T700's price the bigger capacities based on this controller would cost roughly as much:
8TB: ~1300€
16TB: ~2600€
32TB: ~5200€

Personally i doubt we will ever see above 8TB utilized for Gen5 even if the controller itself supports it. The market for 16TB 2500+ price M.2 SSD's is very small.
Especially considering that first gen Gen5 models dont offer that much of a performance increase (perf per W may actually decrease due to higher power consumption on certain workloads) and that anyone can get 4TB Gen4 models for under 250€ now.

Manufacturers should focus on getting 8TB Gen4 to the masses. Currently it's still overpriced and should start from 500€ but starts from 800€ (M.2, yes U.2 2,5" does start from 500€ but U.2 compatibility on desktops is nonexistent).

In my opinion Gen5 is not worth it for the desktop because there are only a few threads doing i/o at the same time. Servers on the other hand can have many independent threads doing i/o.
Benchmarks can have very high throughput because they only move data but don't process that data. Once you need to do something useful with the data the speeds are much lower.
4K 60FPS 32bit per pixel uncompressed video is about 203MB/s and on disk it usually is compressed to much less. Highend camera's can record at 400Mb/s which is just 50MB/s.
So even editing 4k video (which is often quoted as the reason for needing Gen5) doesn't need Gen5. You can do it with Sata600 speeds if you use dedicated disks for in and out streams.

The 4TB Lexar NM790 Gen4 is now selling for about 210 euro and it is fast.
So an 8TB drive could cost less than 500 euro.

One Nand-package usually holds up to 16 die's (of 1Tbit), so 2TB per package.
Samsung has announced a 256TB drive which has 64 packages with 4TB worth of 1Tbit QLC die's per package.
So it's already possible to stack 32 but this is probably neither mainstream or cheap.

Assuming 2TB per package, it's no problem creating 2, 4 & 8TB M2 drives (as long as the controller can handle it).
This should be very cost-effective.

For those looking for large capacity NVMe drives, take a look at the enterprise segment e.g. Samsung PM9A3 with a U2 interface.
Capacities up till 15.36TB are available for reasonable prices. For U.2 drives, adapters to a PCIe slot are available.
And you get powerloss protection and other enterprise features.

I process a lot of data but I don't need this much highspeed storage yet. I have array's of sata-ssd's that serve me just fine.
So I am curious what all those people wanting Gen5 and/or 4 & 8TB drives need them for...

 

[Pizdarenkowitch]

Emerging UK Start-Up Quinas Gets International Prize for its Disruptive Memory Innovations

Quinas Technology, which was originally spun-out of Lancaster University, has just gained a high-profile accolade for its ULTRARAM memory.

www.epdtonthenet.net