14nm AI Chip Hits 520 TFLOPS

zephyr_z9 · x · 2026-07-14

A repost introduces Dongfang Suanxin's DF1000: despite reportedly using a 14nm process, it delivers 520 TFLOPS of BF16 compute. The key lies in DRAM-Logic wafer-level hybrid bonding and 3D vertical stacking, enabling sub-micron interconnects, 6.4 TB/s memory bandwidth, and 900 GB/s scale-up bandwidth.<br><br>The chip is planned for use in TY64 SuperNode and HS128 SuperCluster. The post emphasizes that high-spec compute can still be achieved on older process nodes by leveraging advanced packaging and stacking technologies.

Original post →

More from Infra

Infra channel →