Beijing’s ambitions of joining the world’s top memory-chip makers, and replacing imports with local products in the domestic market, look increasingly at risk as the world’s leading players adopt cutting-edge technologies while China’s ability to catch up is hampered by US sanctions, according to analysts and industry insiders.

This, in turn, could seriously hinder China’s ability to develop large artificial intelligence (AI) models when the country is already reliant on US suppliers such as Nvidia for sanctions-compliant versions of graphics processing units, or GPUs, they said.

Not long ago, China was seen as catching up quickly with international suppliers in the field of advanced 3D NAND flash and DRAM memory chips, but the gap could have widened again in the era of ChatGPT as Yangtze Memory Technologies Corporation (YMTC) and ChangXin Memory Technologies (CXMT) are unable to press ahead with catch up efforts due to the US export restrictions.

In comparison, US-based Micron Technology, which is banned by Chinese authorities from selling certain products to some local clients, last month announced samples of its high-bandwidth memory 3 (HBM3) Gen2 die to support generative AI applications. Samsung Electronics said last month it has completed development of GDDR7 DRAM for AI applications, and it will double down on high-bandwidth memory in 2024 to meet growing demand in AI applications. Meanwhile, rival SK Hynix, which tops the global HBM market with a 50 per cent market share, is also set to double its HBM production next year for AI servers.

Analysts and industry professionals told the South China Morning Post that Micron, Samsung and SK Hynix remain the default choice for many downstream buyers and that Beijing’s restrictions on Micron, which previously had about a 15 per cent market share in China’s server DRAM market, mostly benefited its two South Korean competitors.

“The premier choice of replacement for Micron is still Samsung [and] SK Hynix,” said an executive at a memory controller producer, who asked not to be named due to the sensitivity of the matter.

China’s shortcomings in storage capacity would hinder the country from making the most of its massive data resources in AI development, Ni Guangnan, a Chinese semiconductor expert and an advocate for China’s semiconductor self-sufficiency, told the World Semiconductor Conference Expo in Nanjing last month.

Big Tech Chinese firms from Baidu and Post owner Alibaba Group Holding to Tencent Holdings and JD.com are eager to develop generative AI models amid a heated global AI race, creating a black market for Nvidia GPUs such as the A100 and H100. However, Ni said the importance of data storage is often given less attention in China as the primary focus is on computing power.

Top-tier AI server GPUs have set a new industry standard by primarily using high-bandwidth memory, according to a research note by TrendForce. HBM, a type of advanced DRAM optimised for AI innovation, is only manufactured by Micron, Samsung and SK Hynix.

China has kept its cards close to its chest regarding its latest technologies, leaving the outside world to guess how far it is progressing in Beijing’s self-sufficiency drive. YMTC, for example, has not updated its website news page since October 2022. The company has resorted to local tool suppliers to keep up with its advanced 3D NAND production as strict US export controls have made it harder to go beyond 128 layers, the Post reported previously.

By: DocMemory
Copyright © 2023 CST, Inc. All Rights Reserved