Huawei introduced Peerium, a computing architecture that coordinates up to 1 million AI chips as a single system. The company also announced the Atlas 950 SuperPoD with over 8,000 Ascend 950DT AI accelerator chips. This AI hardware infrastructure aims to train models with up to 40 trillion parameters using UnifiedBus interconnect technology.

Huawei Targets Large-Scale AI Model Training with New Architecture

Huawei has introduced Peerium architecture at its annual CONNECT conference on Sept. 17, a computing system designed to link 1 million AI chips and operate them as a single computer

1

2

. The architecture aims to support training AI models with up to 40 trillion parameters, a scale that significantly exceeds current industry capabilities

1

. For context, competing AI hardware infrastructure from NVIDIA's GB200 NVL72 and Cerebras's CS-3 operates at the trillion-parameter scale rather than tens of trillions

2

. Even Elon Musk's Colossus cluster, described as the world's most powerful AI training system, comprises 200,000 interconnected chips—far fewer than Peerium's target capacity

1

.

Why China Needs This Infrastructure Now

Liao Heng, chief scientist at Huawei's chip design arm HiSilicon, explained the urgent demand driving this development. "Over the next two years, there will be roughly 6 or 7 frontier AI labs in China, and all will aim at training foundation language models of sizes ranging from 10 to 40 trillion parameters," Heng told the Associated Press

1

. This timeline reveals China's aggressive push to compete in the global AI race, with multiple organizations simultaneously pursuing models that dwarf current standards. The infrastructure requirements for such ambitious goals necessitate systems that can coordinate massive numbers of processors efficiently.

Technical Foundation: Three Core Principles

Peerium architecture relies on three fundamental technologies: nested parallelism, unified memory addressing, and peer-to-peer interconnects

1

2

. Nested parallelism breaks workloads into hierarchical groups of AI chips that process tasks simultaneously, similar to how departments and teams coordinate within an organization

1

. Unified memory addressing enables processors to access memory attached to other servers rather than being limited to their own server's memory

2

. Peer-to-peer interconnects allow components to communicate directly without intermediaries, reducing latency and improving coordination

1

.

UnifiedBus Interconnect: The Bandwidth Breakthrough

Huawei rotating chairman Eric Xu identified a critical bottleneck in existing systems. Current interconnect technologies like NVIDIA's combination of NVLink for intra-rack communication and InfiniBand for inter-rack communication provide only 200GB per second of inter-rack communication bandwidth

1

. "With such a low inter-rack communications bandwidth, there is no way you can make one million processors work as one computer," Xu stated

2

. Huawei developed UnifiedBus interconnect technology to address this limitation, delivering up to 800GB per second of inter-rack bandwidth—a fourfold improvement

1

. Xu believes this represents "the path for the future AI era" and expects other players to adopt similar approaches

1

.

Atlas 950 SuperPoD and SuperCluster Deployment

Alongside Peerium architecture, Huawei announced the Atlas 950 SuperPoD, a computing system powered by more than 8,000 Ascend 950DT AI accelerator chips

1

2

. Testing for large-scale AI model training is currently underway on these SuperPods, with results showing more efficient use of available computing power compared to clusters not based on Peerium and UnifiedBus

1

. Huawei expects to begin supplying the SuperPods by early 2027 at the latest

2

. Beyond individual SuperPods, the company is developing SuperCluster systems that link multiple SuperPods to form even larger AI computing networks. One SuperCluster containing 256,000 computing cards is currently being deployed and tested

1

2

.

Source: Live Science

Source: Live Science

© 2026 TheOutpost.AI All rights reserved