Huawei’s Ascend Chip Line: A Game Changer in AI Technology
At the Huawei Connect 2025 event in Shanghai this week, the tech giant unveiled its ambitious plans for the next generation of its Ascend chip series. This innovative development marks a pivotal moment not just for Huawei, but also for the global semiconductor landscape.
Key Highlights from Huawei Connect 2025
During his keynote, Eric Xu, deputy chair of the Huawei board, emphasized that 2025 is a “memorable year” for the company. He reflected on the debut of DeepSeek-R1 earlier this year as a significant turning point. However, Xu acknowledged a pressing challenge: China may lag in semiconductor manufacturing process nodes for the foreseeable future. This admission showcases Huawei’s transparency and strategic approach in the highly competitive tech industry.
Navigating Trade Challenges
In response to ongoing tariffs and trade restrictions, Huawei is not standing still. The company plans to ramp up its infrastructure design and technology. Additionally, it is opting to open-source various software components, including the openPangu foundation AI models and the Mind series SDKs. This move could empower developers and enhance collaboration across the tech community, positioning Huawei as a forward-thinking leader.
The Future of Ascend Chips
Huawei’s latest vision includes three new variants of its Ascend chips: the 950, 960, and 970 series. Each series aims to push the boundaries of computational power and efficiency.
Ascend 950 Series
The Ascend 950PR and 950TO will be manufactured from the same die but will offer distinct enhancements in low-precision data formats, specifically FP8 and MXFP8. The 950 will boast a performance of 1 PFLOP (one quadrillion floating-point calculations per second), while the MXFP8 is rated for 2 PFLOPs. Notably, the memory access has been improved, allowing access down to 128-byte chunks—a significant improvement from the previous 512-byte limitation.
More Read
With an impressive 2 TB/s interconnect bandwidth—2.5 times greater than its predecessor, the Ascend 910C—this chip promises to redefine performance levels. The Ascend 950PR is expected to launch in Q1 2026, while the Ascend 950DT will follow in Q4 2026.
Ascend 960 and 970 Series
Fast forward to Q4 2027, and the Ascend 960 will arrive, featuring double the power, memory access bandwidth, and capacity compared to the 950. This series will also introduce Huawei’s proprietary HiF4 data format, offering enhanced precision over existing FP4 technologies.
At the zenith of performance, the Ascend 970, scheduled for Q4 2028, aims to revolutionize the industry. While specific specs are still under development, Xu hinted that interconnect bandwidth could reach an astonishing 4 TB/s, along with capabilities of 8 PFLOPs in FP4, setting a new standard in chip technology.
SuperPods: The Next Level in Computing
In a bold strategic move, Huawei is introducing SuperPods—clusters of raw compute capabilities designed for hyperscalers. The Atlas 950 SuperPod, featuring the new Ascend 950DT chips, is set to hit the market in Q4 2026. Huawei’s impressive claim? This SuperPod will host 56.8 times more NPUs than NVIDIA’s NVL144 system, promising nearly seven times the processing power.
Even with the upcoming NVL576 from NVIDIA scheduled for release in 2027, Huawei is confident that its SuperPods will maintain superior performance levels.
General Computing with Kunpeng Processors
For general computing needs, Huawei plans to launch two models of its Kunpeng 950 processors by Q1 2026. These processors will feature configurations of 96 cores & 192 threads and 192 cores & 384 threads, catering to diverse computational needs. Additionally, the company will debut what Xu described as the “world’s first general-purpose computing SuperPod,” the Kunpeng 950-based TaiShan 950 SuperPod.
UnifiedBus 2.0: The Future of Connectivity
An exciting development in Huawei’s ecosystem is the introduction of UnifiedBus 2.0, an upgraded interconnection technology based on the existing UnifiedBus 1.0. This new open protocol will facilitate enhanced connectivity within SuperPods and across clusters, ensuring seamless communication between them. The launch of the Atlas 950 SuperCluster, offering 2.5 times more NPUs and 1.3 times more computing power than xAI’s Colossus, showcases the potential impact of this technology.
Come Q4 2027, Huawei plans to unveil the Atlas 960 SuperCluster, which aims to integrate over one million NPUs, delivering a staggering 4 ZFLOPS in FP4.
By ramping up development in semiconductors and enhancing collaboration through open-source initiatives, Huawei is not only solidifying its position as a leader in AI technologies but also reshaping the landscape of global semiconductor manufacturing.
Inspired by: Source
