Sugon 8000 AI Supercomputer at Full Load: 100k‑Card Cluster Processes 500K Jobs

On July 18, during the 2026 World AI Conference, the Sugon 8000 (Dengfeng) supercomputer, labeled the "treasure of the exhibition," released its latest operational data.

In its first week online, the system ran at full capacity, completing more than 150,000 tasks per day, with a single‑day peak of 500,000 tasks.

Sugon 8000 is China’s first fully domestic 100,000‑card AI supercomputing cluster. Beyond demonstrating raw compute scale, it marks the availability of a ten‑million‑card‑class AI infrastructure capable of real scientific and industrial workloads.

The machine follows a native “super‑intelligence fusion” architecture, with co‑design across chips, compute, storage, network, cooling, and services. It supports scientific computing, AI training, and inference on a single platform and is compatible with FP64, FP32, BF16, TF32, FP8, and INT8 precisions.

Stability at the 100k‑card scale stems from engineering advances: a globally unique high‑density cabinet design that raises per‑node compute density by about 20 × compared with other supernodes.

Its self‑developed Scale Fabric IB‑style RDMA network can interconnect up to 10,000 or 100,000 cards, delivering end‑to‑end NIC latency under 1 µs and switch forwarding latency of 260 ns.

Additional technologies—including ParaStor distributed storage, immersion phase‑change liquid cooling, and high‑voltage DC power—work together to sustain massive data flow and long‑term stable operation.

To date, Sugon 8000 has adapted nearly a thousand super‑intelligence‑fusion applications, covering large‑model training, high‑throughput inference, and scientific computing at the ten‑thousand‑card scale.

After connecting to the super‑computing internet, the platform’s daily job count reached 500,000.

Full load in the first week does not imply resource congestion. An intelligent scheduling engine, data‑affinity algorithms, and multi‑modal fusion scheduling policies enable dynamic task routing and on‑demand resource allocation based on job type, data location, and resource status.

From being the exhibition’s "treasure" at WAIC to operating at full load within a week, Sugon 8000 signals that a fully domestic 100k‑card AI supercluster is transitioning from breakthrough to large‑scale practical deployment.