Huawei Connect 2026: Atlas 960E and Peerium Unveiled
Huawei introduces Atlas 960E SuperPoD and Peerium architecture, connecting 1 million chips into a single supercomputer at HUAWEI CONNECT 2026.
Stock photo for illustration only, not from the actual event
- Huawei hosted HUAWEI CONNECT 2026 in Shanghai, focusing on silicon infrastructure for the Agentic AI era.
- Launched Atlas 960E SuperPoD utilizing NPO optical interconnects to deliver up to 8 EFLOPS per rack.
- Unveiled Peerium architecture connecting 1 million chips to work together via UnifiedBus.
- Emphasized an open ecosystem with over 4.16 million developers on the Kunpeng platform.
At HUAWEI CONNECT 2026 in Shanghai, the Techsauce team joined the global technology event to report on cutting-edge industry trends. During the opening keynote, David Wang, Deputy Chairman of the Board and Rotating Chairman of Huawei, took the stage to share his vision under the theme "Advancing the Agentic World, Building a Solid Silicon Foundation," outlining the company's strategy to develop silicon infrastructure for the fully intelligent Agentic AI era.
David Wang pointed out that AI brains are growing faster than any technological revolution in history, projecting foundation model parameter sizes to exceed 100 trillion by 2030, comparable to the number of synapses in the human brain. Meanwhile, inference token usage in China alone has already surpassed 500 trillion tokens per day, necessitating a complete overhaul of traditional IT infrastructure.
Stock photo for illustration only, not from the actual event
To tackle the bottleneck where servers waste over 40% of processing power waiting to transmit data across machines, Huawei developed the SuperPoD concept. By clustering massive numbers of processors via ultra-high-speed data lines, the system unifies all chips to operate seamlessly as a single giant computer, achieving up to 2.75 times higher actual workload efficiency in testing.
A major highlight of the event was the launch of the Atlas 960E SuperPoD, the industry's first supercomputer utilizing Near-Packaged Optics (NPO) technology. Equipped with Hi-ONE optical transceivers delivering 7.2 Terabits per second per device, it significantly reduces heat and saves over 500 kilowatts of electrical power while packing up to 4,096 NPUs into a single cabinet.
"No single company can carry the AI world on its own shoulders."
David Wang, Huawei
Stock photo for illustration only, not from the actual event
Additionally, Huawei introduced the Peerium Computing Architecture, a next-generation processing design allowing all chips to communicate as peers via UnifiedBus. This breakthrough enables a massive scale of 1,000,000 processors to collaborate smoothly. Huawei also highlighted open ecosystem milestones, including over 4.16 million developers on Kunpeng and the openEuler operating system claiming the top spot in China's server market.
Huawei's strategic focus on optical interconnects (NPO) and seamless multi-chip architecture represents a critical milestone in overcoming the physical limitations of traditional silicon hardware, which often suffers from thermal constraints and energy loss during inter-server data transfers. Integrating these technologies enables large data centers to process increasingly complex AI models sustainably and with higher stability moving forward.
Source: Techsauce
Found something wrong in this article? Report an issue with this article
Comments
Leave a Comment