At the Huawei Connect 2023 conference in Shanghai, Huawei’s rotating chairman, William Xu (Wang Tao), unveiled the company’s strategic direction for the intelligent era. The core of Huawei’s AI strategy, Xu emphasized, is compute power, with a persistent focus on monetizing its hardware capabilities.
Key Pillars of Huawei’s AI Strategy
Xu outlined seven key deployments for Huawei’s AI strategy:
- Compute Power as the Core: The fundamental element of Huawei’s AI strategy is compute power, coupled with a commitment to hardware monetization.
- “Supernode + Cluster” Foundation: Huawei aims to build a robust computing power foundation for China and offer new choices to the world through its “Supernode + Cluster” architecture.
- Open and Collaborative Ecosystem: The company is fostering an open and collaborative compute power ecosystem, supporting the native training of mainstream large models and embracing a “hundred-model-dry-state” future.
- Pangu Model Integration: The Pangu large model is being focused on enhancing the intelligence of Huawei’s own products, thereby strengthening their competitive edge.
- Flexible Compute Solutions: Huawei will provide flexible offline and online compute solutions to accelerate industry intelligence.
- Diverse Compute Innovation: The company is driving innovation in diverse computing power, aiming to bring AI to endpoints and vehicles, making intelligence ubiquitous.
- Next-Generation Communication Networks: New communication networks are being built around the concept of “using compute power,” facilitating the transmission of intelligence to individuals, families, and enterprises.
Ascend Supernode Deployment and Future Roadmap
Xu announced that Huawei has already deployed over 1,000 sets of Ascend Supernodes, with the Ascend 950 Supernode now entering mass commercial use. Looking ahead, the Ascend 960 Supernode is set to be the industry’s first to adopt NPO (Non-Persistent Object) technology.
The roadmap for future hardware includes:
- Atlas 960 Liquid-Cooled Supernode: Planned for release in Q3 2027.
- Atlas 860 Air-Cooled Supernode: Scheduled for release in Q2 2027.
Ascend 950 Architecture and Enhancements
During the Huawei Connect 2025 conference, the Ascend 950 chip architecture was revealed. Key advancements include:
- Support for low-precision data formats.
- Enhanced vector computing power.
- A 2.5x increase in interconnect bandwidth.
- Support for Huawei’s proprietary HBM (High Bandwidth Memory).
Specifically, the 950PR variant is designed to improve inference Prefill performance and enhance recommendation service performance, integrating Huawei’s self-developed HBM, HiBL 1.0. The 950DT variant focuses on boosting inference Decode performance and training performance, while also increasing memory capacity and bandwidth.
Previously, the Ascend 960 chip was slated for a Q4 2027 release, indicating Huawei’s continued investment in pushing the boundaries of AI hardware capabilities.
