China is stepping up its chip self-sufficiency push by combining relatively mature chips with new computing architectures in an effort to approach Nvidia’s performance levels, according to a top industry expert. Logic chips made with a 14 nanometer process, which are generally seen as trailing cutting-edge AI processors, could reach a performance comparable with Nvidia’s 4 nm products if they were integrated with high-performance memory and an innovative computing architecture, said Wei Shaojun, Vice President of the China Semiconductor Industry Association (CSIA). He made the remarks at a recent industry event in Shenzhen. Wei, who is also Professor at Tsinghua University’s School of Integrated Circuits, outlined a solution based on “software-defined near-memory computing” that uses three-dimensional (3D) hybrid bonding to stack 14 nm logic chips with 18 nm DRAM. He said this configuration could rival Nvidia’s 4 nm graphics processing units (GPUs) – currently the industry standard for AI training – while significantly cutting costs and power consumption.
His proposal echoes the technological path pursued by Huawei Technologies. Unable to produce 5 nm and more advanced chips at scale because of U.S. export controls, Huawei Founder Ren Zhengfei has argued that China could still achieve state-of-the-art performance by “stacking and clustering” chips rather than competing node for node. Near-memory computing refers to a hardware-software co-design approach that places AI processors close to memory to reduce data movement and improve efficiency. 3D hybrid bonding is an advanced packaging technique that directly bonds chip surfaces to create dense vertical connections in a stack. Wei said his scheme relied on a “fully home-grown supply chain”, underscoring China’s efforts to secure domestic sources of AI computing power as the U.S.-China tech war drags on.
U.S. export controls on advanced chips and chipmaking tools, first introduced in October 2022 and later expanded, cover production of logic chips at 14 nm and below and DRAM at 18 nm and below, constraining China’s access to leading-edge chip manufacturing capacity. Professor Wei also stressed the urgency of reducing China’s reliance on Nvidia hardware. Global AI development remained “deeply bound” to Nvidia’s GPU architecture and its CUDA computing platform, he said, creating what he described as a “triple dependence”, spanning models, architectures and ecosystems.
GPUs, a type of logic chip designed for complex graphics and highly parallel workloads, have become the workhorse for AI training and inference. Chinese players are also betting on other forms of accelerators – such as neural processing units and tensor processing units (TPUs) – to loosen Nvidia’s grip on the AI computing market, the South China Morning Post reports.
