Changxin Achieves HBM3E Breakthrough: Just One Generation Away From Global Leaders

From 动察 Beating AI News: The Information, citing two people familiar with the matter, reports that Changxin Memory has commenced small-batch production of HBM3E, with plans to expand output in 2027. HBM (High Bandwidth Memory) is high-speed memory placed alongside AI chips, tasked with delivering data to the chips at high speeds. Since 2024, the U.S. has restricted exports of HBM2 and higher generations to China, making HBM a key bottleneck for domestic AI chips. Domestic chip developers including Cambricon and Alibaba's T-head are testing Changxin's HBM; if progress continues smoothly, it could be integrated into products as early as next year. HBM3E is the memory generation used by Nvidia's H200 and Blackwell chips. By product generation, Changxin is currently only one generation behind SK hynix, Samsung, and Micron, which are mass-producing HBM4. However, small-batch production is a long way from mature mass production. Changxin's HBM3E still has low yield, and lags behind the three major manufacturers in speed and reliability. SK hynix began mass-producing HBM3E as early as 2024; the three majors have now entered the HBM4 phase and started sampling HBM4E. Changxin has cleared the hurdle of "can we produce it", and the next challenge is to significantly improve yield and output.