Huawei Unveils Next-Gen Ascend Chip Lineup: Ascend 950, 960, and 970 for Superior AI Processing and SuperPod Technology
At the Huawei Connect 2025 event held in Shanghai this week, the global tech titan revealed its plans for future generations of its Ascend chip line.
During his keynote address, Eric Xu, deputy chair of the Huawei board, hailed 2025 as a “memorable year” and highlighted DeepSeek-R1’s debut in January as a significant milestone for the company. He also acknowledged that China might experience a prolonged lag in semiconductor manufacturing process nodes.
In response to trade tensions, Huawei has chosen to accelerate infrastructure design and technology advancement, as well as open-source substantial portions of its software, including the openPangu foundation AI models and the Mind series SDKs.
The tech giant is set to launch three new Ascend chip series: the 950, 960, and 970. The Ascend 950PR and 950TO will be fabricated from the same die and offer additional support for low-precision data formats such as FP8, with the 950 delivering a PFLOP of performance and MXFP8, rated at two PFLOPs.
Improved vector processing and more granular memory access, down to 128 byte chunks from 512 bytes, are also featured in the Ascend 950 chips. These chips will offer a 2 TB/s interconnect bandwidth, 2.5 times greater than the current Ascend 910C. The 950PR is expected to be available in Q1 2026, while the Ascend 950DT is scheduled for release in Q4 2026.
The Ascend 960, set to launch a year later in Q4 2027, will boast double the computing power, memory access bandwidth, memory capacity, and number of interconnect ports compared to the 950. It will support Huawei’s proprietary HiF4 data format, claimed to offer superior precision over other FP4 technologies.
The most advanced chip in the lineup will be the Ascend 970, planned for release in Q4 2028. Xu revealed that the general goal is to significantly boost all of its specifications. The Ascend 970 series is anticipated to offer an interconnect bandwidth of 4TB/s, be capable of 8 PFLOPs of FP4, and come with increased memory capacity.
Huawei’s strategy includes offering hyperscalers raw compute in the form of SuperPoDs, which will debut in Q4 2026 as the Atlas 950 SuperPoD, equipped with the new Ascend 950DT chips. The company claims its first SuperPoD will deliver nearly seven times the processing power compared to NVIDIA’s upcoming NVL144 system.
For general computing, Huawei plans to launch two models of its Kunpeng 950 processors in Q1 2026, featuring 96 cores and 192 threads, and 192 cores and 384 threads in the more powerful model. A “world’s first general-purpose computing SuperPoD,” the Kunpeng 950-based TaiShan 950 SuperPod, will be available in Q1 2026 as well.
The NPU and general computing SuperPoDs will utilize UnifiedBus 2.0, the next iteration of the existing UnifiedBus 1.0. This interconnection technology is used by the Atlas 900 A3 SuperPod, which has seen over 300 installations since its launch in March this year.
UnifiedBus 2.0 will be an open protocol, with technical specifications immediately released to the developer community. UnifiedBus 2.0 will be used internally in the new generations of SuperPods and connect clusters of SuperPods, forming SuperClusters.
The first cluster product is expected to be the Atlas 950 SuperCluster, offering 2.5 times more NPUs and 1.3 times more computing power than xAI’s Colossus, currently the world’s most powerful computing cluster.
In the last quarter of 2027, Huawei intends to launch the Atlas 960 SuperCluster, which will integrate over a million NPUs and deliver 4 ZFLOPS in FP4 (with a ZFLOP representing 10^21 floating point operations per second). “SuperPoDs and SuperClusters powered by UnifiedBus are our answer to surging demand for computing, both today and tomorrow,” Xu concluded.
