Massive SSD density boost is on the horizon thanks to Kioxia's 1,000-layer NAND plan

zohaibahd

Posts: 203   +3
Staff
Forward-looking: Japanese memory maker Kioxia has some encouraging news that could supercharge SSD densities. At the International Workshop on Memory Technology in Seoul, the company laid out an ambitious roadmap to reach a staggering 1,000 layers of 3D NAND flash by 2027. But getting there will be no easy feat.

Kioxia's projections, as covered by the Japanese outlet PC Watch, extrapolate from past trends and improve upon existing NAND cell technology. The company expects NAND die density to hit 100 Gbit/mm² with 1,000 memory cell layers just three years from now. To achieve this, the rate of increase would have to be maintained at 1.33 times per year.

3D NAND layer counts have indeed rapidly increased, from just 24 layers in 2014 to 238 layers in 2022 – a tenfold jump in under a decade. Last year, SK Hynix even showcased a sample of 321-layer 1 Tb TLC 4D NAND chips.

However, scaling to quadruple-digit layer counts is no easy feat. According to storage news site Blocks & Files, achieving higher densities with 3D NAND isn't merely about adding more layers to the chip. Each layer requires an exposed edge to enable connections between memory cells, resulting in a staircase-like die profile. Therefore, as the layer count is increased, the area consumed by the staircase structure grows substantially, offsetting some of the density gains.

To compensate for this, memory makers need to shrink NAND cells both vertically and laterally while transitioning to QLC NAND, packing 4 bits per cell compared to today's TLC technology. Channel resistance and signal noise also turn into growing pains as layers increase.

While Kioxia has plausible solutions for these technical roadblocks, a looming question remains around the economics and financial viability of such an aggressive push.

Kioxia's manufacturing partner, Western Digital, has reportedly expressed concerns about ballooning NAND fab costs outpacing revenue growth. The two companies have already announced their BiCS 8 technology with 218 layers and have discussed BiCS 9 and 10 with up to 400+ layers. However, a 1,000-layer node seems like an ambitious long-term goal that may test WD's appetite for heavy fab investments.

The lengths Kioxia will go to achieve the memory density of its dreams remain to be seen. The manufacturer is currently locked in a race with Samsung, making the 1,000-layered prize a lucrative target. Tough negotiations with Western Digital on the pace and timing of future NAND scaling nodes likely lie ahead.

Permalink to story:

 
PR designed with a specific intention, such as excite investors to prop up valuation, is annoying. Develop the tech first, then announce it. Otherwise, it's just babbling.
 
PR designed with a specific intention, such as excite investors to prop up valuation, is annoying. Develop the tech first, then announce it. Otherwise, it's just babbling.

Let's go pie in the sky venture capitalists . It's a fun game for people already pretty rich, or long shotters.

This is not so much pie in sky, but risk assessment. Can it be done quickly enough, will it be usurped by another tech . how much capital , Return on investment. what is the credentials, character of those involved etc

That is a lot of skill needed
Look at internal investment strategies that have failed - eg Intel, was it their 10nm platform that never arrived.
Boeing with its max plane ( you need to be maxed to the gills to fly on it )

Thing is servers are going to need extra memory at all stages on the CPU/SOC/GPU , on the board Ram/Dimms , SSD and for storage.
However why not just bigger boards with more slots. bandwidth etc

Anyway practically wonder how much heat production it hammered 24/7 by so AI server

Given that imagine the products in a small space, if has 32 TB of very quick access data.
AI could have 2 processing ways like humans, the instant gut reaction and the more considered analysis updated with the 32Tb of models/data
 
This could work if they get the engineering right AND stick with the TLC cell format. They go with QLC and it'll be dead in the water.
 
Back