Huawei designs computing system linking one million artificial intelligence processors
Technology group Huawei unveiled a computing framework connecting one million processors to function as a single system for enterprise artificial intelligence applications
On 17 September 2026, Huawei's Rotating Chairman Eric Xu and Chief Scientist Dr. Liao Heng held a Q&A session at HUAWEI CONNECT to discuss the company's Peerium Computing Architecture for AI and UnifiedBus. At last year's HUAWEI CONNECT, Eric Xu unveiled Ascend chips and their roadmap, including the 950, 960, 970, SuperPoDs and SuperClusters. Today, Huawei announced a SuperPoD powered by the Ascend 950 chip, capable of linking one million artificial intelligence processors.
The architecture, called Peerium, utilizes nested parallelism, unified memory addressing, and peer-to-peer interconnect to achieve strong scaling of processors. This architecture extends the Turing parallelism paradigm with the introduction of Nested BSP (Bulk Synchronous Parallel) and overruns the traditional master-slave design paradigm. The high-speed bus, UnifiedBus, is a key component that scales without limits, connecting CPUs, NPUs, memory, SSDs, interface cards, and switches.
Dr. Liao Heng revealed that the 256,000 computing cards mentioned in his paper are not an arbitrary number. With the goal of training foundation models, which are already in the 5T parameter range, China plans to build 200,000 as a conservative number for the power delivery system in a single autonomous zone and data center campus.
Huawei is actively seeking to encourage Chinese model providers to adopt Ascend chips for training, with the Ascend 950DT SuperPoD expected to start large-scale supply at the end of this year or early next year. Eric Xu expressed that China's AI model providers should aim to reach a level where they can feel the risks from AI development, striking a balance between driving AI development and managing AI risks.
Written by urgent.news from Vietnam Investment Review's reporting — not their text. Machine-written — may contain errors; check the original before relying on it.
