Huawei outlines Ascend 960 roadmap and million-processor AI systems

Huawei has moved up the release schedule for its Ascend 960 chips, with the 960DT due in Q1 2027 and the 960PR in Q3 2027.

Huawei has brought forward parts of its next-generation Ascend processor roadmap as it develops larger computing systems for AI training and inference.

Speaking at Huawei Connect 2026 in Shanghai, David Wang, Huawei's deputy chairman and rotating chairman, said the Ascend 960DT is scheduled for the first quarter of 2027, three quarters earlier than originally planned. The Ascend 960PR is expected in the third quarter of 2027, one quarter ahead of schedule.

"We're evolving our Ascend chip series on a one-generation-a-year cycle," Wang said. Huawei plans to release the Ascend 970 in 2028 and Ascend 980 in 2029, with increases in compute performance, memory capacity, memory bandwidth, and interconnect bandwidth.

Wang said foundation model parameters are approaching 10 trillion and are projected to exceed 100 trillion by 2030. He added that daily inference token consumption in China has reached about 500 trillion and is expected to reach quintillions by 2030.

Huawei scales SuperPoD infrastructure

Huawei is building much of its AI infrastructure around SuperPoDs, which connect multiple computing nodes through high-speed interconnects and unified memory addressing so they can operate as one logical system.

The company said more than 1,000 Atlas 900 A3 SuperPoDs have been deployed, while the Atlas 950 is now in commercial use.

Huawei said communication between processors can account for more than 40% of training time in conventional server clusters. Simulations by its Markov Lab found that a 100,000-NPU cluster built with 4,000-NPU SuperPoDs delivered 2.75 times higher Model FLOPs Utilization than one using eight-NPU servers.

"SuperPoDs are designed to coordinate multiple NPUs through interconnect," Wang said. The Ascend processors form part of an 11-chip portfolio built around Huawei's UnifiedBus architecture, covering computing, networking, storage, and management.

Huawei also announced its Peerium Computing Architecture, which underpins its SuperPoD and SuperCluster systems. The architecture combines nested parallelism, unified memory addressing, and peer-to-peer interconnects to coordinate large numbers of processors as a single computing system.

UnifiedBus serves as the main interconnect layer within the architecture, linking CPUs, NPUs, memory, SSDs, network interface cards, and switches through a common protocol. Huawei said the design is intended to support computing systems that scale to as many as one million processors.

Huawei has also developed Hi-ONE, a near-packaged optical interconnect that provides transmission capacity of up to 7.2 Tbit/s per engine. The company has submitted an implementation agreement on near-packaged optics to the Optical Internetworking Forum.

Huawei is using Ascend 960 processors and Hi-ONE interconnects in the Atlas 960E SuperPoD. A single system can support up to 4,096 NPUs, 8 EFLOPS of FP8 compute performance, and up to one petabyte of high-bandwidth memory, according to Huawei.

The company said the system uses about 5,500 Hi-ONE units instead of roughly 48,000 800G optical modules. Huawei estimates this reduces power consumption by more than 550 kilowatts and provides 99.8% system availability.

Huawei is also upgrading its TaiShan 950 SuperPoD, which supports up to 4,096 nodes and a unified memory pool of up to 256 TB. Huawei said startup speeds for 100,000 sandbox environments were 30 times faster than on traditional servers in its testing.

Huawei has also launched its OceanStor M900 storage cluster for AI inference workloads. The system supports one-hop direct access and provides petabyte-scale key-value cache capacity for workloads involving AI agents and longer context windows.

Huawei combines these systems in a larger SuperCluster architecture, using UnifiedBus to connect Ascend SuperPoDs, Kunpeng-based infrastructure, and KV cache clusters through a common interconnect protocol.

Huawei said an Atlas 950 SuperCluster with 256,000 accelerator cards is already being deployed, while the NPO-based Atlas 960 system is under testing. The Atlas 950 and SuperPoD-based SuperClusters are the first systems Huawei has built around the Peerium Computing Architecture.

Huawei said the SuperCluster can interconnect up to 512,000 NPUs using a two-tier, four-plane Clos architecture. A multi-rail topology can extend the configuration to one million NPUs

Huawei extends its software and device strategy

Huawei is continuing to develop the software ecosystem around its Kunpeng and Ascend hardware.

The company said its Kunpeng ecosystem includes more than 4.16 million developers and 7,200 partners, while openEuler has recorded more than 20 million installations.

Huawei is also continuing open-source development of its Compute Architecture for Neural Networks, or CANN. External contributors now account for 61% of CANN developers, according to the company.

Ascend currently supports more than 90 third-party open-source projects, including PyTorch, Triton, vLLM, and veRL. Huawei said more than 40 AI models have also been pre-trained using Ascend hardware and CANN.

Beyond data centers, Huawei plans four computing platforms covering AI phones, AI PCs, vehicles, and homes.

The company also identified 5G-Advanced, 6G, 10-gigabit optical networks, and multi-tier, low-latency bearer networks as part of its next-generation communications plans for connecting data centers, edge infrastructure, and devices.