A Shanghai-based semiconductor start-up, Dongfang Suanxin, has unveiled the DF1000, a 14nm AI chip that claims to deliver 520 teraflops of BF16 computing performance. According to the company’s specifications, the chip offers 6.4 TB/s of memory bandwidth and 900 GB/s of scale-up bandwidth for inter-chip communication. The company said the DF1000 is ready for mass production, with shipments expected by the end of 2026. Jiemian News adds that the entire DF1000 supply chain is domestically sourced. For comparison, ICSmart notes that the DF1000 delivers approximately 52.6% of the per-chip computing performance of NVIDIA’s Hopper-based 4nm H100/H200, but the report claims the chip reaches 6.4 TB/s of memory bandwidth—nearly twice that of NVIDIA’s H100 and approximately 33% higher than the H200. The chip uses a software-defined architecture combined with 3D-stacked near-memory computing, allowing conventional memory to replace HBM. Dongfang Suanxin plans to launch the DF2000 in 4Q26, with a third-generation DF3000 targeted for late 2027. Founder Wei Shaojun acknowledged that 3D stacking alone has limitations, as stacking multiple silicon layers often lowers manufacturing yields, while limited access to advanced nodes remains the domestic industry’s fundamental constraint.