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

Huawei is on the brink of making waves in the AI landscape with the anticipated launch of the Ascend 910C AI chip, expected to begin large-scale shipments next month. This development signifies a pivotal moment for Chinese tech companies, as they strive for independence from U.S.-produced semiconductors. As restrictions tighten on access to essential hardware, the race for innovative domestic solutions is heating up.

Shifting Dynamics in AI Technology

Recently, Chinese developers faced increased hurdles, particularly regarding access to key Nvidia components. The U.S. government notified Nvidia that its H200 AI chip sales to China now necessitate an export license, prompting local developers to search for alternative solutions capable of supporting extensive training and inference tasks.

The Ascend 910C chip, while not manufactured using the latest process nodes, offers a strategic workaround. Essentially a dual-package iteration of its predecessor, the 910B, this chip boasts two processors, effectively doubling performance and memory capacity. Insiders suggest that its processing power rivals that of Nvidia’s H100.

A Robust Approach to Performance

Huawei is employing an unconventional strategy to boost performance; instead of relying solely on advanced manufacturing processes, it combines multiple chips and high-speed optical interconnects. This novel approach is at the heart of Huawei’s CloudMatrix 384 system—an AI powerhouse designed for training complex models at scale.

• Key Features of the CloudMatrix 384:
  • Comprises 384 Ascend 910C chips across 16 racks, including 12 compute racks and four networking racks.
  • Utilizes optical interconnects, which facilitate high-bandwidth communication, significantly enhancing data transfer rates.

According to SemiAnalysis, the system is equipped with 6,912 800G LPO optical transceivers, creating a robust optical all-to-all mesh network. This architecture positions Huawei to deliver an impressive 300 petaFLOPs of BF16 compute power, surpassing Nvidia’s GB200 NVL72, which registers around 180 BF16 petaFLOPs.

Trade-offs in Efficiency

However, these gains come with drawbacks. Analysts predict that Huawei’s system is approximately 2.3 times less efficient in terms of floating point operations compared to Nvidia’s GB200, alongside lower power efficiency per memory bandwidth unit. Despite these challenges, Huawei’s offerings still lay a solid foundation for training advanced AI models on a large scale.

Domestic Production Challenges

China’s largest chip foundry, SMIC, is reportedly manufacturing crucial components for the 910C utilizing its 7nm N+2 process. While this is a significant step, yield levels remain a concern. Some units of the 910C are said to incorporate chips produced by TSMC for Chinese firm Sophgo, though Huawei has denied using any TSMC-made parts.

The U.S. Commerce Department is currently scrutinizing the relationship between TSMC and Sophgo, particularly following the discovery of a Sophgo-designed chip in Huawei’s previous 910B model. TSMC asserts that it has ceased all dealings with Huawei since 2020, strictly adhering to export regulations.

The Future of Chinese AI Hardware

As of late 2023, Huawei has begun distributing early samples of the 910C to select tech firms and has opened its order books. Consulting firm Albright Stonebridge Group indicates that the Ascend 910C might soon become the preferred choice for Chinese companies aiming to develop substantial AI models or enhance inferencing capabilities amid ongoing export controls on U.S. chips.

While the Huawei Ascend 910C may not compete with Nvidia on power efficiency or cutting-edge technology, it reflects a significant trend: the push for homegrown alternatives to foreign components. Even if this involves employing less advanced methods, the outcome remains noteworthy.

With global demand for AI surging and export restrictions tightening, Huawei’s ability to provide scalable AI hardware domestically could play a transformative role in shaping China’s artificial intelligence trajectory. This development resonates deeply as developers seek stable supply chains and aim to mitigate geopolitical risks.

Embrace the future of AI technology—stay informed and inspired as we navigate these exciting advancements together.