SK Hynix Inc. will adopt the 3-nanometer process, the most advanced foundry technology available, to produce customized HBM4 chips in the second half of 2025, according to semiconductor industry sources on Tuesday.
The South Korean chipmaker had originally planned to manufacture the sixth-generation high-bandwidth memory (HBM4) tailor-made for customers on the 5 nm node. It is teaming up with foundry leader TSMC Co. to develop HBM4.
But it recently shifted to the 3 nm process for HBM4, which it is expected to ship to Nvidia Corp. in the second half of 2025 at the request of its key customers to supply more advanced memory.
SK Hynix is said to unveil a prototype of HBM4 that is vertically stacked on a 3 nm base die as early as March, the sources told The Korea Economic Daily.
Nvidia’s graphic processing unit (GPU) products are currently based on 4 nm HBM chips.
A base die is placed at the bottom of an HBM connected to a GPU and acts as its brain. An HBM stacked on a 3 nm base die is expected to enhance performance by 20-30% compaerd to the HBM4 with a 5 nm base die.
Its shift to a 3 nm foundry process comes as it is going all-in on US Big Tech such as Nvidia Corp., Google LLC and Microsoft Corp. to lessen its reliance on China amid Washington’s clampdown on exports to China
On Monday, Washington announced new semiconductor export controls against China. The latest package includes HBM chips.
SK Hynix commands about half of the world’s HBM market and ships most of its HBM products to Nvidia, the largest AI chip buyer in the world.
Producing HBM4 with a 3 nm base die will further widen its gap with Samsung Electronics Co., which plans to apply its 4 nm foundry process for the sixth-generation HBM, or HBM4.
12 NM FOUNDRY FOR GENERAL-USE HBM4
For general-use HBM4 and HBM4E, SK Hynix will apply a 12 nm process technology in collaboration with TSMC.
The world’s No. 2 memory chipmaker manufactured HBM3E, or fifth-geeration HBM, using its base dies. But for HBM4 chips, it has decided to adopt TSMC’s technology.
TSMC mass-produces 3 nm memory chips embedded in Apple iPhones and MacBooks.
Nano refers to the width of the chip circuit line. The thinner the circuit linewidth, the larger the number of chips can be mounted on a substrate or wafer.
SK Hynix is speeding up the development of HBM4 chips at the request of Nvidia.
Nvidia’s CEO Jensen Huang recently asked SK Group Chairman Chey Tae-won to bring forward the supply of 12-layer HBM4 chips by six months from SK’s schedule of early 2026,Chey said in a keynote speech at the SK AI Summit 2024 last month.
Meanwhile, Tesla Inc. has asked Samsung and SK Hynix to supply HBM4 prototypes for general use, South Korea’s semiconductor industry sources told The Korea Economic Daily last month.
The world’s No. 1 electric vehicle maker is expected to choose one of the two companies as its HBM4 supplier after testing their samples.
By Chae-Yeon Kim and Jeong-Soo Hwang
Why29@hankyung.com
Yeonhee Kim edited this article.
