China's DFSX unveils 14nm DF1000 chip, tops Nvidia H200 on bandwidth

The design trades cutting-edge transistors for smarter memory plumbing, betting that data movement now matters more than raw compute density.

• Shanghai startup Dongfang Suanxin (DFSX) unveiled the DF1000, a 14nm AI accelerator with 6.4TB/s of memory bandwidth and 520 TFLOPS of BF16 compute, edging out Nvidia's H200 (4.8TB/s) despite using a process node roughly a decade behind Nvidia's 4nm chips.
• Founded by Wei Shaojun, DFSX built the chip around what it calls a software-defined architecture and near-memory computing, prioritizing parallelism and data locality over transistor shrinkage, with a supply chain it claims is fully domestic.
• The DF1000 skips scarce HBM memory entirely, instead stacking memory directly atop the compute layer via 3D wafer-level hybrid bonding that replaces microbumps with millions of vertical copper interconnects.
• DFSX's roadmap centers on a TY64 SuperNode built from its next-gen DF2000 chips, claimed to reach 960TB/s of aggregate bandwidth versus 576TB/s for Nvidia's GB200 NVL72 rack — about 67% more, not the literal 2x figure some coverage cites.
• DFSX says DF1000 mass production is underway with shipments expected before the end of 2026, while the more ambitious DF2000 chip is targeted for a Q4 2026 debut.
• The bet is that memory bandwidth, not transistor density, now gates AI inference performance, giving Chinese chipmakers a path to compete by re-engineering packaging on mature, export-control-proof fabrication nodes rather than chasing leading-edge lithography.