SK hynix 发布面向 AI 计算的 3D FeNAND 技术
SK hynix Unveils 3D FeNAND for AI Computing < Semiconductor < 기사본문 - The Elec Inc.
SK hynix 公布基于铁电 3D NAND（3D FeNAND）的模拟存内计算（A-CiM）研究，可在 NAND 闪存内直接执行乘累加（MAC）运算。仿真显示其吞吐量最高达传统非易失性存储器的 20 倍，能效提升 5000 倍。该成果已发表于 IEEE 论文。
正文
当前显示原文，中文全文译文尚未就绪。
가
The body content of the article will be changed to this letter size.
스크롤 이동 상태바
(Source: SK hynix)
SK hynix has developed next-generation memory technology that performs artificial intelligence (AI) computations directly within NAND flash memory.
The company disclosed its research on analog computing-in-memory (A-CiM) using ferroelectric three-dimensional NAND (3D FeNAND) in a newsroom post on Oct. 14. The findings were published in a paper by the Institute of Electrical and Electronics Engineers (IEEE).
A-CiM performs computations directly inside memory using analog signals. It executes multiply-accumulate (MAC) operations, the fundamental building blocks of AI computing, within memory. By reducing data movement, the technology improves processing speed and power efficiency.
3D FeNAND stores data and controls current using the polarization states of ferroelectric materials. Conventional 3D NAND uses a charge-trap nitride (CTN) structure that traps electrons in a nitride layer. Ferroelectric materials retain their internal polarization even after an external electric field is removed.
According to SK hynix, FeNAND can produce multiple current states within a single cell, making it suitable for representing the weights required in AI computations.
SK hynix researchers co-optimized three areas — cells, non-cell circuits and pages — to improve FeNAND’s computing performance. Simulations showed that 3D FeNAND-based A-CiM achieved up to 20 times higher throughput and 5,000 times greater energy efficiency than conventional nonvolatile memory.
At the cell level, current states are distinguished by adjusting the number of programming pulses. Conventional NAND changes cell states by increasing voltage, while FeNAND increases the number of pulses at a constant voltage.
At the non-cell circuit level, researchers reduced wordline resistance. Wordlines select and control memory cells. The SK hynix team introduced air gaps between wordlines.
At the page level, researchers optimized data placement and computation sequences. They applied in-page pipelining (IPP) to efficiently execute AI neural network computations within a single page.
Lee Jun 2jun0504@thelec.kr
My comment