SK hynix Starts Mass Production of World's First 321-High NAND

[btarunr]

SK hynix Inc. announced today that it has started mass production of the world's first triple level cell-based 321-high 4D NAND Flash with 1 Tb capacity. Following its previous record as the industry's first provider of the world's highest 238-layer NAND since June last year, SK hynix has become the world's first supplier of the NAND with more than 300 layers by finding a technological breakthrough for stacking. The company plans to provide the 321-high products to customers from the first half of next year.

Stacking more than 300 layers came into reality as the company successfully adopted the "3 plugs" process technology. Known for its excellent production efficiency, the process electrically connects three plugs through an optimized follow-up process after three times of plug processes are finished. For the process, SK hynix developed a low-stress material, while introducing the technology that automatically corrects alignments among the plugs.



With the adoption of the same development platform from the 238-high NAND on the 321-high product, the company could also improve the productivity by 59%, compared with the previous generation, by minimizing any impacts from a process switch.

The latest product comes with an improvement of 12% in data transfer speed and 13% in reading performance, compared with the previous generation. It also enhances data reading power efficiency by more than 10%.

SK hynix plans to steadily expand the use of the 321-high products by providing them to the nascent AI applications, which require low power and high performance.

Jungdal Choi, Head of NAND Development at SK hynix, said that the latest development brings the company a step closer to the leadership of the AI storage market represented by SSD for AI data centers and on-device AI. "SK hynix is on track to advancing to the Full Stack AI Memory Provider by adding a perfect portfolio in the ultra-high performance NAND space on top of the DRAM business led by HBM."

View at TechPowerUp Main Site

 

[TumbleGeorge]

Strange, so many layers and QLC but still 1TB per chip? What is wrong with math?

 

[Bwaze]

Strange, so many layers and QLC but still 1TB per chip? What is wrong with math?

But it's smaller, it's cheaper for them to make. There's your math!

 

[TumbleGeorge]

But it's smaller, it's cheaper for them to make. There's your math!

From an end customer expecting chips with much increased density and much increased capacity to enable 16TB and 32TB on an M.2 2280 drive, this news does not bring hope. Although hype with some of parameters.

 

[Bwaze]

From an end customer expecting chips with much increased density and much increased capacity to enable 16TB and 32TB on an M.2 2280 drive, this news does not bring hope. Although hype with some of parameters.

In consumer space that's apparently all we want. Same capacity for the last half of decade, and speed increases that only show up when benchmarking.

But that's almost irrelevant now, compared to needs for AI servers. But instead of "trickling down" of tech made for corporate the lines are completely separate now. All we will get are higher prices.

 

[Gungar]

Strange, so many layers and QLC but still 1TB per chip? What is wrong with math?

"Triple level cell" is TLC

 

[R0H1T]

What is wrong with math?

Maths

 

[Wirko]

"Triple level cell" is TLC

Reading through the news once again, I see it's easy to misread 4D as QLC. What's 4D, again? NAND cells losing charge over time?

 

[Bwaze]

Well, you need tine dimension, otherwise you can't change the state of cells.

 

[Readlight]

Now make them socket able.

 

[Gungar]

Reading through the news once again, I see it's easy to misread 4D as QLC. What's 4D, again? NAND cells losing charge over time?

I have no idea, it's probably just 3D with a bit of marketing.

 

[Random_User]

In consumer space that's apparently all we want need. Same capacity for the last half of decade, and speed increases that only show up when benchmarking.

But that's almost irrelevant now, compared to needs whims for AI servers. But instead of "trickling down" of tech made for corporate the lines are completely separate now. All we will get are higher prices.

Fixed. This is quite of reverse way of things. There's absolutely no need in that AI BS. No one has died, yet, as nobody will, without silly AI stuff. The absolutely majority of the back-end/client/consumers do not benefits from it, even for a bit. Only the big companies, for the obvious reasons.

On the other hand, the stagnation in consumer NAND/SSD sector, is just embarassing. The size/speed leaps of HDD technology was much bigger, and in shorter amount of time, than SSDs for a decade (since the first massive appearance). Almost every week the another announcement, from one NAND maker, or other. But the capacity is till like back in 2018, or even earlier.
And that's excluding the fact, that the reliability goes down proportionally the cell capacity.

So people need the cheaper, more reliable storage with more capacity.

 

[Bwaze]

There's absolutely no need in that AI BS. No one has died, yet, as nobody will, without silly AI stuff. The absolutely majority of the back-end/client/consumers do not benefits from it, even for a bit. Only the big companies, for the obvious reasons.

It doesn't matter, the hype is in full swing, even with the reports that AI services aren't generating revenue - it's enough for now that they're generating positive responses in stock market and investors. Who cares about users?

 

[GabrielLP14]

From an end customer expecting chips with much increased density and much increased capacity to enable 16TB and 32TB on an M.2 2280 drive, this news does not bring hope. Although hype with some of parameters.

There are still issues with super high density drives like that on M.2 form factor, like power draw

 

[R-T-B]

But it's smaller, it's cheaper for them to make. There's your math!

Also can use a bigger process to make it more reliable. Its not all about capacity, especially with QLC where endurance already needs all the help it can get.

"Triple level cell" is TLC

Oh it's TLC. Same story but even more endurance. Nice.

 

[Wirko]

Also can use a bigger process to make it more reliable. Its not all about capacity, especially with QLC where endurance already needs all the help it can get.

Oh it's TLC. Same story but even more endurance. Nice.

I don't know but I suspect that endurance needn't be very high for AI use (and what isn't AI these days?). SSDs store models, and models aren't updated often.

 

[R-T-B]

I don't know but I suspect that endurance needn't be very high for AI use (and what isn't AI these days?). SSDs store models, and models aren't updated often.

Still, it's a side effect of using a larger process which also saves them money, so you can bet they did it.

 

[JohH]

They need to treat YMTC seriously. No more conspiracy price fixing with Micron and Samsung. Gotta go all out now or the company will die.

 

[Scrizz]

Reading through the news once again, I see it's easy to misread 4D as QLC. What's 4D, again? NAND cells losing charge over time?

All NAND cells lose charge over time....

 

[Wirko]

They need to treat YMTC seriously. No more conspiracy price fixing with Micron and Samsung. Gotta go all out now or the company will die.

Price increases tend to self-synchronise without human intervention across world regions, political systems, etc. Here's the NM790 in euros:

 

[JohH]

without human intervention

Incorrect. They conspire all the time and have been convicted of it in the past. And they likely did it again. But YMTC won't play by their games because it is state-financed.