Huawei accelerates Ascend 960DT to 2027

Huawei has announced that the launch of its Ascend 960DT AI chip has been accelerated to the first quarter of 2027. This announcement was made during the Huawei Connect conference on September 17. The Chinese tech company is positioning itself directly against Nvidia, which currently sets the standard in AI hardware. With the Ascend 960DT, Huawei aims to combine hundreds of thousands of AI chips into one system to significantly increase computing capacity, both for training and execution.

According to Huawei, the Ascend 960 chips will be available earlier with the new timeline, and computing power will double each year. This acceleration was announced by David Wang, the acting chairman of Huawei. Huawei Connect provides the platform for the company to share its plans for chips and cloud solutions, and the acceleration of chip development underscores their ambition to play an active role in the competitive AI landscape.

The foundation of Huawei's strategy is the Peerium Computing Architecture. This architecture utilizes UnifiedBus technology, allowing multiple AI chips to work together effectively. This enables Huawei to develop larger AI systems for model training and inference, with the first implementations based on this architecture, such as the Atlas 950 SuperPoD and the Atlas 950 SuperCluster. However, the recent announcement raises questions about the scalability of the systems, as previous reports regarding chip configurations have not been confirmed.

Huawei's progress is also influenced by political pressure, with speed becoming a strategic factor in the development of AI technology and chips. Eric Xu, rotating chairman of Huawei, emphasized the importance of the Peerium architecture for building advanced AI systems. With the release of the Ascend 960DT in 2027, Huawei clearly positions itself in the competitive arena alongside Nvidia and other AI technology providers, without providing details on pricing or delivery timelines.

Read the full article from AI Insider.