Huawei has advanced its launch schedule for two Ascend 960 AI chips, targeting the 960DT for the first quarter of 2027 and the 960PR for the third. The schedules are three quarters and one quarter earlier, respectively, than the company’s original roadmap.
Rotating chairman David Wang outlined the plans at Huawei Connect in Shanghai. The company plans a new Ascend generation each year, with the 970 in 2028 and the 980 in 2029. Those dates are planned releases, rather than announcements that the chips are shipping.
Huawei also unveiled the Atlas 960E SuperPoD, using Ascend 960 chips and near-packaged optical interconnects. Its stated specifications include up to 4,096 neural processing units, eight exaflops of FP8 computing performance and one petabyte of high-bandwidth memory. The company says its Hi-ONE optical engine provides 7.2 terabits per second of transmission capacity.
The new hardware sits within an 11-chip portfolio built around UnifiedBus, Huawei’s interconnect technology for computing, storage and management. Wang said Huawei has shipped more than 1,000 smaller linked computing systems to more than 370 customers. The company is developing larger systems in which many processors share information and operate together.
In a separate technical announcement, Huawei named its system design Peerium Computing Architecture. It combines nested parallelism, unified memory addressing and peer interconnect, with UnifiedBus connecting CPUs, AI processors, memory, solid-state drives, network interface cards and switches through a common protocol.
Huawei describes the architecture as extending computing to the million-processor scale. Its current deployment account is narrower: an Atlas 950 SuperCluster with 256,000 cards is being deployed, while the near-packaged-optics Atlas 960 system is under testing. The million-processor figure describes the architecture’s stated capacity; the deployment cited in the announcement uses 256,000 cards.
The Atlas 950 SuperPoD and clusters built from it are the first generation using Peerium. Huawei says the common bus provides peer connections across computing, storage and networking, avoiding separate communication paths that require protocol conversion. These system specifications and deployment descriptions are Huawei’s own claims.
DeepSeek is separately reported to be planning an Inner Mongolia cluster with at least 160,000 Ascend 950DT accelerators. That proposed installation is intended for inference, according to the reporting, despite the 950DT’s training-oriented design. The order concerns the existing 950 generation, rather than the newly scheduled 960 chips.
Optical interconnect suppliers are also moving products into manufacturing. Tower Semiconductor and NewPhotonics have begun volume shipments of laser-integrated optical engines for 800G and 1.6T connections, with a separate 6.4T chipset planned for 2027.
Huawei Connect runs through September 19 at the Shanghai World Expo Exhibition and Convention Center and Shanghai Expo Center. The 960PR availability target is the third quarter of 2027.
–
By the Control Plane Editorial Team