Chinese chip startup Dongfang Suanxin has unveiled a processor architecture designed to challenge NVIDIA while sidestepping the advanced fabrication nodes restricted under US export controls, betting on software-defined computing and 3D-stacked memory to close the performance gap.
The company’s flagship DF1000 processor, built on a 14-nanometer process, delivers 520 teraflops of BF16 performance with 6.4 TB/s memory bandwidth and 900 GB/s inter-chip bandwidth. It is ready for mass production with shipments expected by the end of 2026. The roadmap targets the NVIDIA H200 with the DF2000 later this year and the B300-level performance with the DF3000 in 2027.
“We have to forge a path of our own. That path cannot be about passively catching up within a framework set by others,” said Wei Shaojun, founder of Dongfang Suanxin. “We need independent architecture, original technology, a self-sustaining ecosystem and a secure, controllable supply chain.”
The architectural bet rests on two pillars. Software-defined computing allows the processor to dynamically reconfigure its compute and data-flow resources to match different workloads, reducing reliance on raw transistor scaling. 3D-stacked near-memory architecture places memory vertically closer to the processing cores, cutting data travel distance for lower latency and reduced energy consumption, performance gains that do not depend on a smaller fabrication node.
The company, founded in Shanghai approximately two years ago, is backed by state investment funds as well as venture capital from Xiaomi, JD.com, and Yunfeng Capital. It has also developed a supporting ecosystem including the Dianfeng accelerator module, TY64 supernode, QY100 integrated computing appliance, and the CAAP open software stack that supports mainstream AI frameworks.
The approach reflects a broader shift among Chinese chip companies toward architectural innovation as an alternative to the shrinking process nodes controlled by US export policy. The challenge acknowledged by the company is that stacking multiple silicon layers can reduce manufacturing yields, and limited domestic access to advanced nodes remains the industry’s biggest obstacle to higher absolute performance.
Sources: China’s new chip startup uses 3D design to challenge NVIDIA despite US curbs (Interesting Engineering, July 2026)