Huawei Unveils Cutting-Edge Ascend Chips to Revolutionize Global Supercomputing Clusters

Chinese technology titan Huawei has unveiled ambitious plans to elevate its Ascend chip line during the Huawei Connect 2025 event in vibrant Shanghai. In a riveting keynote address, Eric Xu, the deputy chair of the Huawei board, reflected on 2025 as a pivotal year, marking significant advancements for the company. The introduction of the DeepSeek-R1 earlier this year serves as a cornerstone for Huawei’s future, despite acknowledging China’s ongoing challenges in the semiconductor manufacturing realm.

In response to trade barriers and tariffs that have impacted the industry, Huawei is not just adapting but innovating. The company is actively enhancing its infrastructure design and technological capabilities while making strides to open-source various software, including the openPangu foundation AI models and the Mind series SDKs.

The Next Generation of Ascend Chips

Huawei plans to launch three new series of the Ascend chip: the 950, 960, and 970.

• The Ascend 950PR and 950TO will share the same die, enhancing capabilities with low-precision data formats like FP8. The 950 series is set to deliver an impressive 1 PFLOP of performance, while the MXFP8 will achieve 2 PFLOPs. To put that into perspective, a PFLOP stands for a staggering one thousand trillion floating-point calculations per second.

• Improvements in vector processing and memory access will allow for finer granularity, reducing memory chunks from 512 bytes to 128 bytes.

• Notably, the Ascend 950 chips will provide 2 TB/s of interconnect bandwidth—an astonishing 2.5 times that of the current Ascend 910C. Expect to see the 950PR in Q1 2026 and the 950DT by Q4 2026.

Looking ahead, the Ascend 960, debuting in Q4 2027, promises to double computing power, memory access bandwidth, capacity, and the number of interconnect ports compared to the 950. It will prominently support Huawei’s HiF4 data format for enhanced precision over conventional technologies.

The pièce de résistance will be the Ascend 970, anticipated in Q4 2028. Xu hinted that specifications are still in the works, but the goal is to significantly elevate its capabilities. Early projections suggest an interconnect bandwidth of 4 TB/s and an astonishing performance capacity of 8 PFLOPs in FP4.

Introducing SuperPods of NPUs

Huawei’s bold strategy includes delivering hyperscaler clusters of raw computing power, branded as SuperPods. The initial model, the Atlas 950 SuperPod, is set to roll out in Q4 2026, featuring the new Ascend 950DT chips.

Competitor NVIDIA is also launching its NVL144 system, a similar SuperPod format, anticipated in mid to late 2026. Huawei asserts that its SuperPod will outperform NVIDIA’s version, offering 56.8 times more NPUs compared to GPUs in the NVL144, delivering nearly seven times the processing power. Even with NVIDIA’s upcoming NVL576 set for 2027, Huawei is confident that the Atlas 950 SuperPod will remain a formidable contender.

Advancements in General Computing Chips

In the realm of general computing, Huawei is gearing up to unveil two variations of its Kunpeng 950 processors in Q1 2026. The models will feature configurations of 96 cores and 192 threads, alongside a faster variant housing 192 cores and 384 threads. Additionally, there’s the launch of the Kunpeng 950-based TaiShan 950 SuperPod, touted as the world’s first general-purpose computing SuperPod.

A New Era with Open-source Connectivity Protocols

The innovative UnifiedBus 2.0 will interconnect both NPU and general computing SuperPods. This next-gen protocol builds on the existing UnifiedBus 1.0, which has already seen successful implementations with over 300 installations since its debut in March.

One of the exciting aspects of UnifiedBus 2.0 is its open protocol nature, providing tech specifications instantly to the developer community. This interconnection technology will facilitate internal connections in the upcoming generations of SuperPods, forming multidimensional SuperClusters.

The first product in this lineup will be the Atlas 950 SuperCluster, expected to deliver 2.5 times more NPUs and 1.3 times greater computing power compared to xAI’s Colossus, currently the reigning champion in computing clusters. Looking a bit further, in late 2027, Huawei plans to launch the Atlas 960 SuperCluster, integrating over a million NPUs and promising an eye-popping 4 ZFLOPS in FP4 performance.

As Eric Xu emphasized, “SuperPods and SuperClusters powered by UnifiedBus are our answer to surging demand for computing, both today and tomorrow.”

Take inspiration from these innovation-driven advancements that define the future! Explore how you can harness the power of technology to elevate your endeavors. Join us in this exciting journey toward unparalleled computing capabilities!