At its Connect event, Huawei outlined its goal of developing domestically produced AI chips and announced a comprehensive AI chip roadmap extending until 2028. With this roadmap, the company aims to compete with global companies like Nvidia by focusing specifically on domestic technology.
Huawei is adopting a new chip strategy
The first significant product on the roadmap is the Ascend 950PR, the successor to the previously introduced Ascend 910C. This new chip utilizes Huawei’s in-house developed HBM technology. The 950PR achieves 1 PFLOPS of computing power in FP8 and 2 PFLOPS in FP4, supporting low-precision data formats. The chip’s interconnect bandwidth is 2 TB/s.
Huawei will utilize HiBL 1.0 HBM technology in the 950PR, offering a capacity of 128 GB and a bandwidth of 1.6 TB/s. The company also plans to offer 144GB of capacity and 4TB/s of bandwidth with its second-generation HiZQ 2.0 HBM memory. The Ascend 950PR is positioned as an “inference” chip designed specifically for caching and recommendation algorithms.
Huawei’s roadmap includes not only inference but also training-focused chips. The Ascend 950DT, scheduled for release in the fourth quarter of 2026, will be training-focused and offer higher bandwidth and memory capacity with HiZQ 2.0 HBM technology.
The company also has ambitious targets for 2027 and 2028. The Ascend 960, due to be released in the fourth quarter of 2027, will feature 2.2TB/s of interconnect bandwidth, 288GB of HBM memory, and 9.6TB/s of memory bandwidth.
This chip’s computing performance was announced as 2 PFLOPS at FP8 and 4 PFLOPS at FP4. Finally, the Ascend 970, due to be introduced in 2028, will offer a comprehensive solution to China’s AI needs with significant improvements in memory and computing power.
{{user}} {{datetime}}
{{text}}