Huawei Launches Next-Gen Ascend Chips and SuperPoDs at Connect 2025

Prime Highlights

Huawei introduced Ascend 950, 960, and 970 chips, with the flagship 970 expected to deliver 8 PFLOPs and 4 TB/s bandwidth by 2028.
The company also launched Atlas SuperPoDs and SuperClusters, claiming performance levels that surpass NVIDIA’s NVL series.

Key Facts

Huawei’s Atlas 960 SuperClusterwill feature over 1 million NPUs and achieve 4 ZFLOPS FP4 performance.
To address interconnect challenges, Huawei rolled out UnifiedBus 2.0, an open protocol for scalable connections across large AI systems.

Background

At the Huawei CONNECT 2025, the company announced new Ascend chips, high-performing SuperPoDs, and SuperClusters to increase AI infrastructure across the globe.

Huawei Deputy Chairman Eric Xu emphasized the importance of computing power as the basis of AI, pointing to the fact that Ascend chips remain at the center of the company’s strategy. In spite of the difficulties in the semiconductor production process, Huawei said it will continue to innovate by advancing chip design, releasing new processors, and making key software open-source.

The company revealed three new chip series: Ascend 950, 960, and 970. The Ascend 950, arriving in Q1 2026, has an interconnect bandwidth of 2 TB /s and a performance of up to 2 PFLOPs. The 960, projected to arrive in 2027, will boost the computing power and memory capacity by twofold, and the flagship, the 970, is projected to attain 8 PFLOPs and 4 TB/s bandwidth when it arrives in 2028.

Huawei also introduced the Atlas 950 and 960 SuperPoDs, housing 8,192 and 15,488 NPUs, respectively. These AI-optimized systems will, according to Xu, outperform current competitors, including NVIDIA’s NVL series. Expanding further, Huawei unveiled the Atlas 950 SuperCluster with over 500,000 NPUs and the Atlas 960 SuperCluster with more than one million NPUs, offering up to 4 ZFLOPS FP4 performance.

For general-purpose computing, Huawei released TaiShan 950 SuperPoD, the first of its kind in the world that could support enterprise workloads as well as AI training.

To overcome interconnect limits, Huawei launched UnifiedBus 2.0 as an open protocol, enabling scalable connections across SuperPoDs and SuperClusters. Xu said this innovation will ensure Huawei can meet rising global demand for computing power.