Huawei Initiates Mass Shipments of Ascend 910C Despite US Restrictions

Huawei's Ascend 910C AI Chip Set for Mass Shipment

Huawei is gearing up to ship its Ascend 910C AI chip on a large scale as early as next month, sources close to the company reveal. Although a few units have already been dispatched, this move signifies a crucial advancement for Chinese companies in their quest for homegrown alternatives to US-made semiconductors.

Context of the Launch

The timing of this rollout is critical, as Chinese developers are grappling with tightened restrictions on Nvidia hardware. The US government recently mandated that Nvidia's H20 AI chip sales to China require an export license, pushing local developers to seek solutions capable of handling extensive training and inference tasks.

Technical Specifications of the Ascend 910C

The Ascend 910C, while not crafted with the latest process nodes, offers a practical solution. It's essentially an enhanced version of the 910B, featuring a dual-package design with two processors that boost both performance and memory. Insiders claim its performance is on par with Nvidia's H100.

Instead of chasing the latest manufacturing technology, Huawei has opted for a more robust approach. By integrating multiple chips and high-speed optical interconnects, they aim to amplify the system's performance. This strategy is key to their CloudMatrix 384 system, a comprehensive rack-scale AI platform designed for training large-scale models.

CloudMatrix 384: A Closer Look

The CloudMatrix 384 system incorporates 384 Ascend 910C chips spread across 16 racks—12 for computing and 4 for networking. Unlike traditional copper-based systems, Huawei's design leverages optical interconnects, which facilitate high-bandwidth communication across the system. According to SemiAnalysis, it includes 6,912 800G LPO optical transceivers, creating an optical all-to-all mesh network.

This configuration allows the CloudMatrix 384 to deliver around 300 petaFLOPs of BF16 compute power, surpassing Nvidia's GB200 NVL72 system, which tops out at about 180 BF16 petaFLOPs. Additionally, the CloudMatrix boasts superior memory bandwidth and capacity, offering over twice the bandwidth and more than 3.6 times the high-bandwidth memory (HBM) capacity of its Nvidia counterpart.

Performance Trade-offs

However, the Huawei system isn't without its compromises. It's estimated to be 2.3 times less efficient per floating point operation compared to Nvidia's GB200 and has lower power efficiency per unit of memory bandwidth and capacity. Despite these drawbacks, Huawei's solution provides the necessary infrastructure for training advanced AI models at scale.

Manufacturing and Supply Chain

China's leading chip manufacturer, SMIC, is reportedly producing key components for the 910C using its 7nm N+2 process. Yet, yield concerns persist, and some units of the 910C reportedly contain chips from TSMC, produced for Chinese firm Sophgo. Huawei, however, denies using TSMC components.

The US Commerce Department is currently probing the ties between TSMC and Sophgo, following the discovery of a Sophgo-designed chip in Huawei's earlier 910B processor. TSMC insists it hasn't supplied Huawei since 2020 and remains compliant with export regulations.

Market Reception and Future Outlook

Late in 2023, Huawei started distributing early samples of the 910C to select tech firms and opened its order books. According to consulting firm Albright Stonebridge Group, the chip is poised to become the preferred choice for Chinese companies developing large AI models or deploying inference capabilities, especially amidst ongoing US export controls on semiconductors.

Although the Ascend 910C might not compete with Nvidia in terms of power efficiency or advanced process technology, it represents a significant shift. Chinese tech firms are increasingly focusing on creating domestic alternatives, even if it means employing less advanced techniques to achieve similar results.

With the global demand for AI solutions on the rise and export restrictions becoming more stringent, Huawei's capability to provide a scalable AI hardware solution domestically could play a pivotal role in shaping China's AI landscape. This development is particularly crucial as developers seek to secure their long-term supply chains and mitigate geopolitical risks.