Huawei has announced its Peerium Computing Architecture, a groundbreaking innovation that allows one million processors to function as a single computer. This architecture, unveiled at HUAWEI CONNECT on September 17, 2026, is designed to meet the demands of the AI era. Eric Xu, Huawei's Rotating Chairman, and Dr. Liao Heng, Chief Scientist of HiSilicon, discussed the architecture and its applications with international media outlets. Peerium Computing Architecture enables strong scaling to the million-processor level through nested parallelism, unified memory addressing, and peer-to-peer interconnect.
The Peerium architecture extends the parallelism paradigm of Turing with the introduction of Nested BSP (Nested Bulk Synchronous Parallel). This architecture also extends the von Neumann single-machine architecture and overruns the master-slave design paradigm that has prevailed for decades. To support this architecture, Huawei invented a key interconnect technology – UnifiedBus. UnifiedBus is a high-speed bus that uses an open protocol to scale without a limit to connect CPUs, NPUs, memory, SSDs, interface cards, and switches.
Huawei has built the Atlas 950 SuperPoD, a product based on the Peerium Computing Architecture. A SuperCluster with 256,000 computing cards is currently being deployed and tested. According to Dr. Liao Heng, the number of 256,000 cards is not arbitrary, but rather a result of the hardware infrastructure needed to support the training of foundation models, which today are already in the 5T parameter range.
The demand for large-scale AI computing infrastructure is driven by the need to train foundation language models of sizes ranging from 10 to 40 trillion parameters. In China, most data centers have a national plan for power grid design, which limits the scale of infrastructure to 200,000 cards. Huawei is currently testing and deploying the Atlas 950 SuperPoD, with large-scale supply expected to start at the end of this year or early next year.
Huawei is having extensive dialogues and discussions with AI model providers to encourage them to use Ascend chips for training. The testing results of using the Ascend 950DT for training have been promising, and Eric Xu believes that starting from next year, a lot of AI model training will be based on SuperPoDs that use the Ascend 950DT. However, the supply capacity will depend on the production volume.
Regarding recent calls by leading AI model providers in the US to slow down AI development, Eric Xu noted that the level and cadence of AI development in China and the US differ. While AI models in China are largely open-source, mainstream model providers in the US have more computing power and may be more aware of the risks associated with AI development. Huawei aims to strike a balance between driving AI development and managing AI risks.
Huawei currently has no plans to expand the Atlas 950 SuperPoD to international markets in a fully-fledged way, as it struggles to meet demand in China. However, the company is doing some testing and providing limited supplies to countries with strong demand. Huawei's focus is on meeting the growing demand for AI computing infrastructure in China, where the company is playing a significant role in promoting chip self-sufficiency.
Key points
- Huawei introduces Peerium Computing Architecture, enabling one million processors to work as one computer for AI applications.
- The Atlas 950 SuperPoD, built on Peerium, is being deployed and tested with 256,000 computing cards.
- Huawei aims to balance AI development and risk management, with a focus on meeting growing demand in China.
