China's hybrid-bonded AI accelerators could rival Nvidia's Blackwell GPUs — top semiconductor expert hints at 'fully controllable domestic solution'

At an industry event, Wei Shaojun discussed China's potential to develop AI accelerators using 14nm logic chiplets and 18nm-based DRAMs that could rival Nvidia's Blackwell processors. He proposed a 'fully controllable domestic solution' combining 14nm logic with 18nm DRAM using 3D hybrid bonding for improved performance efficiency. While this solution is hypothetical, Wei believes it could offer 120 TFLOPS performance with 60W power consumption, surpassing Intel's Xeon CPUs in efficiency. The key technologies for this accelerator include 3D hybrid bonding and near-memory computing, aiming to compete with Nvidia's CUDA platform and advance AI hardware competitiveness in China.