520TFLOPS computing power, 6.4TB/s memory access bandwidth, and Oriental Computing Core’s first large computing power chip DF1000 released

Zhidongxi Author | ZeR0 Editor | Mo Ying Zhidongzhi reported on July 13 that today, Shanghai AI chip startup Dongfang Core released the software-defined near-memory computing 3D AI chip DF1000. The DF1000 chip is the first AI chip in China to adopt DRAM-LOGIC Wafer-level hybrid bonding 3D vertical packaging. It adopts software-defined chip technology and 14nm process technology. Based on the national supply chain, it achieves a computing power of 520TFLOPS@Bf16, a memory access bandwidth of 6.4TB/s, and a scale-up bandwidth of 900GB/s. Through vertical stacking and integration of computing and storage layers, the chip compresses the interconnect spacing from tens of microns to sub-micron levels, resulting in an order of magnitude increase in interconnect density and bandwidth density. It fundamentally solves the bottlenecks such as "storage wall", "bandwidth wall" and "power consumption wall" that computing power systems have long faced. At the same time, this chip can realize a full closed loop from chip design, wafer manufacturing, 3D advanced packaging to sealing and testing under the conditions of domestic mature chip manufacturing processes. It does not rely on cutting-edge manufacturing processes and is not afraid of external constraints. The supply chain is stable and controllable and can be iterated sustainably. It provides a practical technical path for domestic high-end chips to break through process dependence. The DF1000 chip has completed complete tape-out verification and achieved full-function and stable operation of a 128-card large-scale cluster. Its real-life scenario implementation capability has reached the industry's advanced level. On this basis, Oriental Computing Core continues to iterate products and is expected to launch DF2000 in the fourth quarter of this year. All performance parameters will be doubled compared to DF1000, with the advantages of high computing power utilization and large access bandwidth. It is expected to launch DF3000 in the fourth quarter of next year, and all aspects of the indicators will be doubled on the basis of DF2000. Today, Oriental Compute Core simultaneously launched a full-stack product matrix based on DF1000, including Pinnacle DF1000 accelerator card, Tuoyu TY64 super node, Qingyuan QY100 server, and Huisuan HS512 smart computing cluster. The entire product system can be adapted to various complex computing power scenarios such as artificial intelligence, Internet, finance, and supercomputing. "DF1000