Hynix launches 321-layer NAND | Hacker News Reader

Hynix launches 321-layer NAND

electronicsweekly.com

125 pointsWaitWaitWha46 comments

What's interesting is that these devices don't need 321+ litho steps; the vertical layers are all defined with deposition. Lithography step count isn't layer dependent it seems.

https://youtu.be/ANHzVOiUwGI

https://thememoryguy.com/3d-nands-impact-on-the-equipment-ma...

When I was in school studying NAND devices (2004-2010) we were quite apprehensive at the long term quantum stability of 4-layer devices.

This (the past 20 years of improvement) is an incredible feat of engineering.

This is 321 physical layers of silicon in an IC, not 321 charge levels.

QLC flash - with 16 charge levels, for four bits per cell - is pretty common nowadays, but that's as far as it goes so far. And stability is indeed a concern; modern flash devices rely heavily on error correction.

Memory like HBM uses stacking of silicon layers, but NAND flash isn't layers of silicon, it's other materials for capturing and holding charge with still just one silicon plane it's all built on top of I believe.

What does 'layer' mean in this context? I'm only familiar with planar style logic process nodes which have maybe up to 20 layers (and way more lithography steps to manufacture those layers), but I am completely ignorant of how the term is used for a flash process node.

How many layers are needed for each physical cell? Is it 1,2, or a lot more? Is this effectively 321 physical TLC cells stacked vertically and some planar style logic at the bottom of the stack.

Also, where do multiple pieces of silicon factor into this - I assume we might be up to 16 silicon dies deep with through-silicon-vias, which would mean a cross section of a package could actually have 5000 layers - that sounds crazy!

Probably done with 3 separate litho/etch layers, where they etch and process in groups of 110 or so.

Each of those layers can have a cell, so if you have a tlc device at a 100nm pitch, you have a density of 321*3/(1e-4)^2 bits/mm, or about 1e11bits/mm2.

Fun reference: atomic density is 1atom/.5nm, so 1/5e-7^2, or 4e12/mm2 ish.

Not too far away.

Amazing, I had no idea how far things had diverged between logic and flash since the move to 3D.

https://borecraft.com/files/Comparison_Current_NAND.pdf (from 2019) has some of the cross-sections I was looking for - and that only goes up to 96 layers!

If I am reading this correctly, this is still the same 1Tb per die but with 321 layers, meaning up to 2TB / package. The package should now be under 100mm2.

This would hopefully bring down the price of 4TB and 8TB SSD in the near future.

Isn't the controller a big problem in drives over 2tb?

Are you referring to amount of channels supported by controllers ?

I just want smaller SPI flash for embedded :( it's been over 10 years since there's been improvement in that space

WLCSP-8 is pretty damn small already, at ~1.5mm square. Hard to get much smaller.

But they aren't high capacity. As far as I've seen, we've been stuck on the same XY size 512 Mb dies for over a decade. Even now, Infineon is claiming they have a series that'll go up to 4 Gb but is still at the standard 2 Gb maximum. NOR hasn't gotten any denser in forever.

What's the use case for a NOR flash that large? Even at 2 Gbit, you're probably better off with something optimized for density like eMMC.

The use is that there's often (in my field) no space for an 11.5x13 eMMC. There are some that are slightly smaller, but as you brought up the wlcsp-8, there's nothing like eMMC/high capacity NAND density scaled down. If I had the bits/mm^2 of even NAND 5 years ago, I'd be a happy camper. But that's life.

There's QSPI NAND parts available; they're just annoying to use.

https://www.digikey.com/en/products/detail/winbond-electroni...

https://www.digikey.com/en/products/detail/alliance-memory-i...

There's also 9x8mm eMMC. The big issue with shrinking it further is that it tends to be a module with a separate controller doing lots of things to make the memory reasonable to use.

Yeah those are what I'm looking at, but even then we've been at 8Gb for years. Manufacturers only want SLC NAND in these for valid reasons and I guess the market isn't pushing for now. The 9x8 is useful but the 3.3v that eMMC wants means I can't power off a single cell li-ion without a boost. It's all a nightmare. Trust me I've looked for solutions, unless you know of any silver bullets that came out recently.

And as you surely know, I usually can't boot from NAND (due to the aforementioned annoyance) so I'd have a boot flash and a storage flash and that's unideal.

I'll note though that the controllers are small. You can RE the die size of a common eMMC<->NAND controller and it's much smaller than 9x8. I won't share which because I honestly don't remember if we got an NDA in place but considering they all stack dies in there anyway, I don't really see that as the size driver.

A lot of MCUs can boot from XIP QSPI/OSPI NAND. quite a feat of compatibility engineering - they made the NAND page size match QSPI transfer sizes commonly used to populate caches, so instead of bit level reads, the flash supports only cacheline level reads, which is usually what you need for XIP anyway.

It's too bad not every embedded device is an MCU :/

Nearly anything that can boot from XIP flash can, plenty of MPUs too, also many Intel chips.

Yeah I'm meaning below MCU boot capability, not above

Could you give an example?

There’s no money in it. Embedded doesn’t pay compared to phones/tablets. So companies are putting their money into that

I'm fully aware of why there's been no improvement, it just sucks for me.

I left embedded when I realized everything, at least in my neck of the woods, was going to end up being cramming phone parts into things that were not phones. The writing has been on the wall for a while now.

Even now as a consumer I can see the stagnation. It's the same parts year after year. Or you become a phone. You have my sympathy.

Are you saying that partly why every device is a “smart” device? Because it’s cheaper to fit components with connectivity already built in so you might as well use it?

At embedded world 2024 they pushed IoT for other reasons. They wanted us to do it for security and updates. No more telling users to put a bin on a flash drive and plug it into a hidden port under the coffee maker or something idk, just be Internet connected. I wouldn't ever say it's cheaper materialsb (though ESP32s are very cheap) or upfront labor, but I'm sure it's cheaper on the support side for bug fixes and stuff. And then they can sell your data, too. Never forget that

That's what this is looking like tbh. I guess I'm just hoping for a miracle. Maybe chiplets will save us.

Where’s the point where you figure out how to stack chiplets perpendicular to a backplane instead of doing lithography 300 times on the same chip?

That's kind of what we're doing already, although the stacking is parallel, not perpendicular. A lot of the innovation is in how the dies are tied together, cf. https://www.anandtech.com/show/9520/toshiba-brings-throughsi...

This is an interesting testament to manufacturing reliability - that they can go to so many layers and still achieve good yield.

As others here have pointed out -- you don't need to do lithography 300 times. The big breakthrough of 3D NAND is depositing alternating layers of films to build most of the layers.

This is 3 stacks of >100 layers.

> triple level cell-based 4D memory

What does 4D memory mean?

Nothing - just marketing. It's their 2nd gen Periphery Under Cell (PUC) device.

It’s marketing speak. 3D flash (stacked chips) with the control circuits stacked underneath instead of to the side. So it’s one louder.

https://www.tomshardware.com/news/sk_hynix-debuts-4d_nand,37...

Bigger number = more betterer

It'd have to be something involving time... maybe have each cell be a delay line? Or a resonant cavity, and store multiple bits at different modes? Not sure how either could be made small enough to be worth it though...

Wow. What's the yield like? Are some bits bad and bypassed during testing?

> Are some bits bad and bypassed during testing?

Always. All digital storage media depends on error correcting codes and sector-remapping these days.

Cookie permission dialog is the worst I have encountered in months

Thank god for ublock origin filters, i have all the optional filter lists, I never see those things. EVER.

I noticed that as well.

I'm not sure what the rules are, but I had to disable a surprising number of "legitimate interests" related to advertising.

Title should probably be "Hynix launches 321-layer NAND RAM".

I think it might be flash memory (i.e. non-volatile, unlike RAM), but this must be so obvious to readers of this publication that it's never said explicitly.

insider info: all the top talent at Samsung left for SK Hynix after government stepped and forced DEI on Samsung leading to unqualified managers ruining Samsung's culture of innovation and rewarding experimentation.

When did this happen?

Reply on news.ycombinator.com