With the Atlas 960E system, Huawei is establishing a new era of ultra-efficient optical interconnectivity for training artificial intelligence models with tens of trillions of parameters at the highest standards of sustainability and energy conservation.
- Launch of the revolutionary Atlas 960E system
- HiOne optical interconnect architecture and performance specifications
- TaiShan system and context memory storage cloud
- Cloud services and agent platforms for enterprises
- Open-source support and ecosystem building
- Frequently asked questions
Launch of the revolutionary Atlas 960E system
Huawei leveraged its annual global conference, Huawei Connect, held in Shanghai, China, to launch what it described as the world’s first supercomputing "SuperPod" system relying on "near-chip optics" technology, abbreviated as NPO. The company also unveiled a broad portfolio of AI-oriented cloud infrastructure services and digital agent platforms for the business sector, in a strategic move confirming its position as an integrated, comprehensive, and independent technical alternative in the global race for intelligent computing and big data processing.
HiOne optical interconnect architecture and performance specifications
David Wang, Huawei’s executive director and chairman of the investment review board, delivered the conference keynote on September 17 under the title: "Driving the world of intelligent agents forward and building a solid silicon foundation." He unveiled the Atlas 960E SuperPod system specifically designed to accelerate training and inference generation processes for giant artificial intelligence models reaching up to 10 trillion parameters. This giant system centers around the innovative HiOne optical engine, which Huawei presents as the industry’s first mass-produced commercial optical solution equipped with an internally integrated light source, boasting a massive data transmission and transport capacity of up to 7.2 terabits per second.
By integrating 5,500 units of HiOne engines, the Atlas 960E system completely eliminates 48,000 traditional 800-gigabit optical units that were previously required, helping reduce electric power consumption by more than 550 kilowatts while achieving an operational availability and stability rate of up to 99.8%. A single system can scale and connect up to 4,096 neural processing units (NPUs) via the Unified Bus full-optical interconnect network, delivering a phenomenal computing performance of 8 exaflops at FP8 calculation precision and 16 exaflops when using FP4 precision. Furthermore, multiple systems can be linked and aggregated into a giant super cluster supporting up to 512,000 NPUs, with exceptional scalability reaching up to one million processing units when applying multi-path network topologies.
TaiShan system and context memory storage cloud
Wang’s announcements did not stop there; he also revealed the TaiShan 950 SuperPod system, which applies the same advanced optical architecture to general-purpose multi-purpose computing environments, offering an ultra-large unified memory pool reaching up to 256 terabytes. In parallel, the company introduced the OceanStor M900 context memory storage pool, which provides a petabyte-scale distributed key-value cache (KV Cache) memory designed to support the operation of large models with extended context windows.
Cloud services and agent platforms for enterprises
On September 18, Peter Zhou, president of Huawei Cloud, showcased a parallel set of advanced digital services and products. The announcements included the launch of the AI Cloud Clusters service, which enables the continuity of model training cycles for more than 40 consecutive days without interruption, featuring self-recovery from software and hardware failures in just 10 minutes, and achieving 20% higher text token generation productivity compared to the previous generation. Commercial rollout of this cloud service will begin in the Chinese market on September 30, and it will become available in international markets outside China on November 30.
Huawei Cloud also introduced its new Agents-as-a-Service platform, which enables developers and engineers to invoke and use models from various providers and platforms without needing to independently deploy each model. The company expanded the scope of its AgentArts enterprise digital agent platform, which currently serves over 100 major institutions and enterprises, connecting with more than 5,000 digital assets and tools supporting the Model Context Protocol (MCP) and over 1,000 specialized assets dedicated to specific industrial sectors. AgentArts will be available outside China by December 30.
Open-source support and ecosystem building
At the conclusion of his remarks, David Wang emphasized Huawei’s steadfast commitment to developing open-source software for its comprehensive computing platform, CANN, stressing that the company seeks to transition from the stage of "usable systems" to "user-friendly systems" through enhanced community developer collaboration. The OpenGeoWin agent community has surpassed 50,000 stars and more than 3.29 million global downloads. Additionally, Huawei Cloud added two new technical domains—smart government and AI hardware—to its industrial AI foundry platform, which now hosts over 1,000 industrial assets distributed across seven specialized sector regions.
Wang concluded his speech with an ambitious message to the industry: "We will continue to provide full, localized support for training leading global foundation models on Ascend chips, and we enthusiastically embrace a broad and diverse spectrum of models and computing applications. Together, let us lead the world toward the era of pervasive intelligence."
Frequently asked questions
Question: What is the core technical advantage of Huawei’s Atlas 960E SuperPod system?
Answer: Its reliance on near-chip optics with HiOne engines, which eliminates thousands of optical units and reduces power consumption.
Question: What is the computing processing capacity offered by a single Atlas system?
Answer: It offers a capacity of 8 exaflops at FP8 precision, reaching 16 exaflops at FP4 precision, and supports up to 4,096 neural processors.
Question: When will the new cloud services and agent platforms be available for global markets outside China?
Answer: The cloud clusters service will be available internationally on November 30, followed by the AgentArts platform on December 30 of this year.
