Huawei Launches Mass Shipments of Ascend 910C Processor Despite US Trade Restrictions

Huawei’s Ascend 910C AI Chip: A Game-Changer for China’s Tech Landscape

Huawei is gearing up for a significant milestone in the tech industry with the anticipated large-scale shipments of its Ascend 910C AI chip, set to commence next month. This development is crucial for Chinese companies that are actively seeking domestic alternatives to US-made semiconductors, particularly in light of recent restrictions impacting access to Nvidia’s hardware.

The Context of Huawei’s AI Chip Launch

The backdrop of this launch is the tightening grip of US export controls, which have made it increasingly challenging for Chinese developers to source high-performance AI chips from American manufacturers. Recently, the US government mandated that Nvidia must secure an export license for its H20 AI chip sales to China, leaving local developers scrambling for viable solutions to support their large-scale AI training and inference workloads.

Technical Specifications of the Ascend 910C

While the Ascend 910C is not manufactured using the most cutting-edge process nodes, it offers a compelling alternative through its innovative design. The chip operates as a dual-package version of its predecessor, the 910B, effectively doubling the performance and memory capabilities. Early reports suggest that the 910C’s performance is comparable to Nvidia’s H100 chip, highlighting its potential as a formidable competitor in the AI hardware space.

A Brute-Force Approach to Performance

Huawei’s strategy with the Ascend 910C involves a brute-force approach to enhance performance. Instead of relying solely on advanced manufacturing techniques, the company has opted to combine multiple chips and utilize high-speed optical interconnects. This methodology is integral to Huawei’s CloudMatrix 384 system, an expansive AI platform designed for training large-scale models.

Architecture of the CloudMatrix 384 System

The CloudMatrix 384 system is a remarkable feat of engineering, featuring 384 Ascend 910C chips distributed across 16 racks—12 dedicated to computing and four for networking. This architecture leverages optical interconnects instead of traditional copper, facilitating high-bandwidth communication between system components. According to analysis from SemiAnalysis, the system is equipped with 6,912 800G LPO optical transceivers, enabling an all-to-all mesh network configuration.

More Read

Exploring Namibia’s Hydrogen Potential and Improving AI Evaluation Methods

Clio Acquires vLex for $1 Billion: Major Move in Legal Software and Data Solutions

Overcoming AI Data Challenges: What Businesses Need to Know

Why Intempus Advocates for Robots to Emulate Human Physiological States

4 Key Charts Highlighting the Current Landscape of AI and Energy

Performance Metrics and Comparisons

In terms of computational power, the CloudMatrix 384 is capable of delivering around 300 petaFLOPs of BF16 compute performance, surpassing Nvidia’s GB200 NVL72 system, which achieves approximately 180 BF16 petaFLOPs. Furthermore, the CloudMatrix boasts substantial advantages in memory bandwidth and capacity, offering more than double the bandwidth and over 3.6 times the high-bandwidth memory (HBM) capacity compared to its competitors.

Efficiency Considerations

Despite these impressive gains, there are trade-offs to consider. Analysts predict that Huawei’s system may be 2.3 times less efficient per floating-point operation than Nvidia’s GB200. Additionally, the power efficiency per unit of memory bandwidth and capacity is also lower. Nonetheless, the Huawei CloudMatrix 384 provides the necessary infrastructure for training advanced AI models at scale, making it a viable option for organizations in need of robust AI capabilities.

Manufacturing and Supply Chain Dynamics

The production of the Ascend 910C involves China’s largest chip foundry, SMIC, which is utilizing its 7nm N+2 process for key components. However, concerns regarding yield levels persist, and there are reports that some 910C units incorporate chips produced by TSMC for the Chinese firm Sophgo. Huawei has publicly denied the use of TSMC-made parts, amid ongoing investigations by the US Commerce Department into the relationship between TSMC and Sophgo.

Market Implications and Future Prospects

As of late 2023, Huawei has started distributing early samples of the 910C to select technology firms and has opened its order books. Consulting experts from Albright Stonebridge Group suggest that the Ascend 910C is poised to become the preferred choice for Chinese companies focused on building large AI models or enhancing their inference capabilities, particularly in light of the ongoing export controls affecting US-made chips.

A Broader Trend in Domestic Semiconductor Development

While the Huawei Ascend 910C may not match Nvidia in terms of power efficiency or advanced process technology, its emergence indicates a broader trend within the Chinese tech sector. Companies are increasingly investing in homegrown alternatives to foreign components, even if it involves utilizing less sophisticated manufacturing methods to achieve comparable outcomes.

As global demand for AI technology continues to rise, Huawei’s ability to provide a scalable AI hardware solution domestically could play a pivotal role in shaping the future of artificial intelligence in China. This strategic move not only supports the nation’s tech aspirations but also mitigates risks associated with geopolitical tensions and supply chain vulnerabilities.

Discover More About AI Innovations

For those interested in exploring the forefront of AI and big data, events like the AI & Big Data Expo in Amsterdam, California, and London offer a unique opportunity to learn from industry leaders. This comprehensive event is co-located with other significant conferences, including the Intelligent Automation Conference and Cyber Security & Cloud Expo, making it a must-attend for anyone invested in the future of technology.

Inspired by: Source