购买与服务热线:400-810-0466

服务邮箱:Support@sugon.com

首页存储 scaleX40 supernode + DeepLink Fabric, leading a new paradigm for China-developed computing-power interconnect

scaleX40 supernode + DeepLink Fabric, leading a new paradigm for China-developed computing-power interconnect


Recently, At the WAIC 2026 Pujiang Ecosystem Forum, Shanghai Artificial Intelligence Laboratory, together with Sugon and other China-developed computing-power industry-chain partners, jointly released the HPC-AI fusion computing-power architecture for scientific intelligence, DeepLink Fabric.As The architecture core hardware foundationSugonthe world's first cable-free chassis supernodescaleX40laying a critical hardware foundation for China-developed computing-power interconnect with hard-core strength.


20260728-1.jpg



Breaking the static-topology bottleneck, scaleX40supports programmable architecture


Traditional computing-power clusters use a static, fixed topology, with one interconnect path serving all tasks, so performance compromises are always present.DeepLink Fabric introduces a topology-programmable design, It uses dynamic interconnect so different computing tasks can switch freely across heterogeneous computing power, shortening data-transfer paths and lowering communication overhead to release overall computing-power potential, meeting the complex mixed-precision computing-power needs of scientific intelligence scenarios.


ShanghaiWang Hui, a young scientist at the AI laboratory, explained that it is like building an entirely new “urban multi-level road network,” sending large data onto “wide lanes,” small data onto “narrow lanes,” and urgent data onto “fast lanes,” with the network actively matching traffic demand rather than forcing vehicles to follow existing routesand then moves from “task-adapted interconnect” to “interconnect-adapted tasks.”


Sugon's fully in-house scaleX40 high-density supernodeWith surging computing power, ultra-high reliability, and extreme interconnect, for this transformative architecture achieves scaled commercial useprovides core hardware vehicle.A single scaleX40 node integrates 40 GPUs, with aggregate computing power exceeding 28PFLOPS (FP8 precision), total HBM memoryover 5TB, with memory bandwidth over 80TB/ssystem reliability reaches 99.99%, safeguarding large-scale stable operation; It adopts an orthogonal, cable-free first-level interconnect architecture in which compute nodes and switch nodes plug directly together, eliminating signal loss and operations risk caused by cables at the source.


In addition, scaleX40 dynamic intelligent scheduling and millisecond-level fault self-healing, further improving the new architecture's system availability.relying on scaleX40's built-in high-speed dynamic interconnect capability, the system can switch interconnect posture in real time according to task traits such as high-throughput training, low-latency inference, and neighborhood communication for scientific computing, enabling dynamic partitioning and real-time multipath schedulingscaleX40 supernodealso It offers low-latency forwarding and millisecond-level fault self-healing, providing a commercially ready hardware premise for full-chain latency exposure and fully unleashing IOD/Chiplet short-reach interconnect, CPO/NPO optoelectronic synergy and other advanced interconnect value.



20260728-2.jpg



Open standards and industrial collaboration, scaleX40 leading new computing-power infrastructure


forum, Sugonalso together with Shanghai Artificial Intelligence Laboratory and industry-chain partnersjointly Jointly building public collaboration standards around four directions: standardized topology description, cross-domain resource scheduling, joint interconnect-performance validation, and convenient heterogeneous computing-power access.As a core contributing author of the Technical System White Paper on Supernodes, Sugon contributed a complete set of engineering practices for 10,000-card cluster networking and high-density supernodes, and participated as a domestic computing-power hardware representative in the IEEE international working group on scientific intelligence, promoting inclusion of the supernode approach in global computing-power interconnect specifications.


This marks China-developed computing-power infrastructure moving from single-point hardware breakthroughs to system-level upgrades. Going forward, Sugon will using scaleX40 supernode as the starting pointjoining hands with more industry-chain partners, using a full-stack independent computing-power foundation to support the comprehensive landing of scientific intelligence and industrial intelligence.#Sugon 


# Sugon News

相关产品和解决方案

联系我们

售后服务

严正声明