China exploring ways to make its own AI memory chips despite US sanctions - South China Morning Post

techsooper.blogspot.com

China is exploring ways to produce its own high-bandwidth memory (HBM), the next-generation of memory chips tailored for artificial intelligence (AI) processors, as it pushes ahead with a semiconductor self-sufficiency drive amid US sanctions, industry sources said.

While it will be an uphill battle to catch up with global leaders like SK Hynix, Samsung Electronics and Micron Technology given the impact of Washington’s sanctions, the Chinese government has determined that the country must become self-sufficient in HBMs even though it may take years, they added.

Industry sources, who are familiar with the matter but declined to be named due to the sensitivity of the subject, said China’s top dynamic random access memory (DRAM) maker, ChangXin Memory Technologies (CXMT), is the country’s best hope for HBMs, but it may take up to four years to bring products to the market.

China quietly poaches chip talent from elite US universities

If CXMT or other Chinese chip makers decide to go ahead, they will be restricted to using less advanced technologies to manufacture the powerful DRAM chips that are in hot demand around the world, the sources said.

Hefei-based CXMT, which reportedly is planning an initial public offering this year based on a valuation of US$14.5 billion, did not reply to a request for comment.

SK Hynix, considered a leader in the technology with a 50 per cent global market share, developed HBM3 in October 2021 and entered mass production in June 2022. The Korean company noted in its promotional materials that HBM technology was a “prerequisite for Levels 4 and 5 of driving automation in autonomous vehicles”.

Demand for HBM chips is expected to grow by almost 60 per cent in 2023, as they are the preferred solution for overcoming memory transfer speed restrictions due to bandwidth limitations, according to a report by tech consultancy TrendForce.

SK Hynix announced last week that it had successfully developed HBM3E, the next-generation of high-end DRAMs for AI applications, and was providing samples to a customer for product performance evaluations. Mass production is scheduled for the first half of 2024, with clients such as US chip firms AMD and Nvidia reportedly lining up for the new product.

China’s AI push feels the heat from US ban on venture capital funding

Nvidia set a new industry standard by using HBM chips to speed up data transfers between graphic processing units (GPUs) and memory stacks, according to TrendForce. Its much-sought after H100 graphics processor unit (GPU) features a HBM3 memory system, delivering 3-terabytes per second of memory bandwidth, according to Nvidia.

HBM vertically stacks memory chips, like floors in a skyscraper, effectively shortening the distance information has to travel. These towers of memory dies connect to a CPU or GPU through an ultra-fast interconnect called the “interposer.”

Besides SK Hynix, other dominant players in HBM are Samsung Electronics and US-based Micron Technology.

Industry insiders said that despite their high performance, the production of HMB chips does not necessarily require cutting-edge lithography technology such as extreme ultraviolet (EUV) tools, making it possible for China to produce its own versions even without the latest equipment. Due to the need to integrate several chips vertically, high-density packaging technologies such as through-silicon-via (TSV) are required, but China has relatively advanced players in this field, such as Jiangsu Changjiang Electronics Technology.

“I wouldn’t be surprised if CXMT is making its own efforts in HMB,” said an executive at a memory chip controller company. He said CXMT can make DRAMs at the 17 to 19-nanometre node, a number of generations behind global peers who are at sub-10-nm, in a similar way to how China is producing 28-nm logic chips.

Adblock test (Why?)