Huawei Atlas 960E: 8 EFLOPS, 4,096 NPU Specs & Roadmap

Introduction

At HUAWEI CONNECT 2026 in Shanghai, Huawei announced the Atlas 960E SuperPoD, a system designed for large-scale AI workloads. The Atlas 960E SuperPoD is intended for organizations requiring significant compute resources for advanced AI development, including language models and agentic applications. According to Huawei, the system is engineered to handle training and inference for models approaching 10 trillion parameters.

Huawei Atlas 960E SuperPoD and Ascend 960: Architecture and Roadmap Overview

Atlas 960E SuperPoD Architecture and Technical Specifications

Huawei reports that the Atlas 960E SuperPoD can scale to 4,096 NPUs (Neural Processing Units) within a single system. This is achieved through the UnifiedBus interconnect and Hi-ONE near-packaged optics. The system delivers 8 exaFLOPS (EFLOPS) of FP8 compute performance and can be configured with up to one petabyte of high-bandwidth memory (HBM). Huawei also claims the system can provide 16 EFLOPS at FP4 precision.

A key architectural feature is the integration of 5,500 Hi-ONE units, which, according to Huawei, eliminates the need for 48,000 800G optical modules. This design reduces power consumption by over 550 kilowatts. Huawei states this approach doubles the system’s fault-free operating time and achieves 99.8% system availability.

The Hi-ONE engine is described as a mass-produced near-packaged-optics solution with 7.2 Tbit/s transmission capacity and a built-in light source, supporting high-density, high-bandwidth connections between NPUs.

Ascend 960 Series Roadmap

Huawei has outlined a roadmap for the Ascend 960 series. The company states that development of the Ascend 960 has exceeded expectations, with performance doubling as planned. The next major release, Ascend 960DT, is scheduled for the first quarter of 2027—three quarters ahead of the original roadmap. The Ascend 960PR is planned for the third quarter of 2027, one quarter earlier than previously scheduled.

Looking ahead, Huawei plans to release the Ascend 970 in 2028 and the Ascend 980 in 2029, maintaining a one-generation-per-year cadence. Each generation is expected to bring increases in compute performance, memory bandwidth, memory capacity, and interconnect bandwidth.

System Design and Ecosystem

Huawei’s strategy centers on large-scale computing systems and interconnect architectures, aiming to scale out system-level resources. The company has described an 11-chip UnifiedBus portfolio for SuperPoDs and SuperClusters, focusing on modular, scalable interconnect solutions. Huawei is also building its Ascend and CANN ecosystem as an alternative for Chinese developers, providing a domestic platform for AI development and deployment.

Release Timeline and Future Developments

Both Ascend 960DT and 960PR are scheduled for release in 2027, ahead of the original roadmap. Huawei plans to release additional generations of chips through 2029. The company also unveiled new AI computing and chip technologies at its annual conference in Shanghai.