At HUAWEI CONNECT on September 17, 2026, Huawei's Rotating Chairman Eric Xu and Dr. Liao Heng, Chief Scientist of HiSilicon, detailed the company's new Peerium Computing Architecture and UnifiedBus (UB) interconnect during a roundtable with international media. The architecture, announced alongside the Ascend 950 chip and Atlas 950 SuperPoD, is designed to make one million processors work as a single computer, a leap beyond conventional SuperPoD designs that typically link only thousands or tens of thousands of processors.
Peerium achieves strong scaling to the million-processor level through nested parallelism, unified memory addressing, and peer-to-peer interconnect. It extends the parallelism paradigm of Turing with Nested Bulk Synchronous Parallel (Nested BSP) and overruns the von Neumann single-machine architecture, breaking from the master-slave design paradigm that has prevailed for decades. Because existing industry technologies could not support this architecture, Huawei invented UnifiedBus, a high-speed bus using an open protocol that scales without limit to connect CPUs, NPUs, memory, SSDs, interface cards, and switches. With UnifiedBus, peer-to-peer interconnect is achieved across compute, storage, and networking, fundamentally displacing traditional master-slave architectures.
The Atlas 950 SuperPoD is the first product built on Peerium. A SuperCluster with 256,000 computing cards is currently being deployed and tested. Dr. Liao explained that the 256,000-card scale is not arbitrary: it supports training of foundation models already in the 5-trillion-parameter range, and over the next two years, six or seven frontier AI labs in China are expected to aim for models of 10 to 40 trillion parameters. He added that 200,000 cards is a relatively conservative number given power-grid designs for single autonomous zones and data center campuses in China.
Eric Xu said the Ascend 950PR, primarily for AI inference, has launched with limited supply, while SuperPoDs based on the Ascend 950DT for AI training are under testing and will see large-scale supply starting at the end of this year or early next year. He noted that training results with the 950DT are promising and that many AI model training workloads will likely run on these SuperPoDs next year. He stressed that supply capacity, not persuasion, will determine adoption, and urged the value chain to expand faster.
On AI risk, Xu observed that Chinese models are largely open-source and transparent, while US providers have more computing power and may have a keener sense of risk. He said Chinese model providers need to accelerate to reach a level where they also feel those risks, and he called for a balance between driving AI development and managing AI risks, with the bottom line being AI for good, not AI for evil.
Regarding international expansion, Xu said Huawei lacks enough capacity to fully satisfy demand in China, so it has no fully-fledged plans to expand abroad. Some testing and limited supply are occurring in countries with strong demand. He also claimed that, based on Huawei's data, Ascend has surpassed NVIDIA in China, though he noted that collecting market-share data is difficult. Xu framed China's push for chip self-sufficiency as inevitable, saying that even if its chips are less advanced, assured supply matters for a country of 1.4 billion people and an industrial powerhouse.
The full Q&A session is available via the https://www.apmultimedianewsroom.com/multimedia-newsroom/huawei-pioneers-a-new-computing-architecture-for-the-ai-era-making-one-million-processors-work-as-one-computer-2. The announcement was distributed through Media OutReach Newswire.
