China tries for AI ICs

HiSilicon’s Ascend 910B is used  in Huawei’s public cloud infrastructure as well as being sold outside,  and  Baidu ordered over a thousand of the chips to build approximately 200 AI servers.

While Chinese company iFlytek, in partnership with Huawei, released the “Gemini Star Program,” a hardware and software integrated device for exclusive enterprise LLMs, equipped with Ascend 910B.

Baidu developed its first self-researched ASIC AI chip, Kunlunxin, in early 2020, with its second generation scheduled for mass production in 2021 and the third expected to launch in 2024.

Post-2023, Baidu aims to use Huawei’s Ascend 910B acceleration chips and expand the use of Kunlunxin chips for its AI infrastructure.

After Alibaba’s acquisition of CPU IP supplier Zhongtian Micro Systems in April 2018 and the establishment of T-Head Semiconductor in September of the same year, the company began developing its own ASIC AI chips, including the Hanguang 800.

TrendForce reports that T-Head’s initial ASIC chips were co-designed with external companies like GUC.

However, after 2023, Alibaba is expected to increasingly leverage its internal resources to enhance the independent design capabilities of its next-gen ASIC chips, primarily for Alibaba Cloud’s AI infrastructure.

The U.S. sanctions cover both software and hardware. Notably, in October 2023, the U.S. Department of Commerce added companies like Biren and Moore Threads to the Entity List.

Additionally, regulations governing advanced manufacturing processes, such as logic ICs with processes finer than 16nm, DRAM with processes finer than 18nm, and NAND Flash with more than 128 layers designated for export to China, were introduced.

These measures have extended the review criteria for AI chip hardware design beyond total processing performance to include performance density requirements, thereby complicating the supply of high-end AI chips from leading manufacturers like NVIDIA and AMD.

Beyond the 2023 U.S. sanctions, the latter half of 2022 saw significant restrictions on EDA semiconductor design software tools, particularly affecting the design of advanced processes like Samsung’s 3nm or TSMC’s 2nm technologies.

Although the mainstream market chips, such as NVIDIA’s A100 and AMD’s MI200, are based on the 6/7nm process, and upcoming models like NVIDIA H100 and AMD MI300 series are expected to shift to 4/5nm processes by 2024, TrendForce forecasts that, despite EDA restrictions not having an immediate significant impact in the short term, they will pose long-term challenges for China in adopting more advanced processes and in the development of next-gen, higher-performance HPC or AI chips