Dongfang Suanxin Unveils HBM-Free 14nm AI Chip, Claims Higher Memory Bandwidth Than Nvidia H200东方算芯发布无HBM 14纳米AI芯片，宣称内存带宽超越英伟达H200

Shenzhen-based Dongfang Suanxin this week unveiled an AI accelerator built on a 14-nanometer process node — a specification that SMIC can manufacture at volume — that carries no high-bandwidth memory. The company claims the design nonetheless delivers higher memory bandwidth than Nvidia's H200, an assertion that, if validated, would challenge the conventional assumption that HBM is a prerequisite for competitive inference performance.

The strategic significance is twofold. First, HBM remains one of the hardest components for China to source domestically given U.S. and allied export controls on SK Hynix and Micron advanced memory. A design that eliminates that dependency entirely removes a key supply-chain vulnerability. Second, 14nm is a node where SMIC has established production, meaning domestic fabrication at scale is plausible without ASML's EUV tooling.

The chip targets inference workloads — consistent with where Chinese enterprise AI spending is concentrating as foundation models move into production deployment. Independent benchmark verification has not yet been published.