Huawei Commences Large-Scale Shipping of Ascend 910C Despite US Restrictions

Sources suggest that Huawei is poised to commence widespread distribution of its Ascend 910C AI chip as soon as next month.

Although a few units have already been sent out, moving to mass production represents a significant milestone for Chinese companies aiming to replace American semiconductor technology with domestic solutions.

This development aligns with a period when Chinese developers are experiencing stricter limitations on Nvidia hardware access. Recently, the US government notified Nvidia that exporting its H20 AI chip to China now necessitates a special license. Consequently, developers in China are exploring alternative options capable of managing large-scale training and computation tasks.

The Ascend 910C, while not utilizing the most advanced manufacturing techniques, offers a strategic alternative. It acts as an augmented version of the older 910B, essentially doubling performance and memory by integrating two processors. Insiders familiar with the chip report performance on par with Nvidia's H100.

Instead of leaning on state-of-the-art production, Huawei employs a powerful strategy of combining several chips with fast optical connections to elevate performance levels. This technique underpins Huawei's CloudMatrix 384 system, an extensive AI platform designed for training complex models.

The CloudMatrix 384 uses 384 Ascend 910C chips distributed across 16 racks, which include 12 for computing and four for networking. In contrast with copper-based systems, Huawei relies on optical links, facilitating high-speed data exchange throughout the system. According to SemiAnalysis, this setup uses 6,912 800G LPO optical transceivers to create a comprehensive optical mesh network.

This configuration enables Huawei's setup to achieve around 300 petaFLOPs of BF16 computing power, surpassing Nvidia's GB200 NVL72, which hits about 180 BF16 petaFLOPs. The CloudMatrix claims additional benefits with greater memory bandwidth and capacity, boasting more than twice the bandwidth and over 3.6 times the high-bandwidth memory (HBM) capacity.

Nonetheless, these capabilities come with trade-offs. The Huawei system is anticipated to be 2.3 times less efficient per floating-point operation compared to Nvidia's GB200, and it features lower efficiency per memory bandwidth unit and capacity. Despite this, Huawei's platform still furnishes the essential infrastructure for extensive AI model training.

Semi-Conductor Manufacturing International Corporation (SMIC), China's leading chip manufacturer, is reportedly responsible for producing some of the vital components for the 910C using its 7nm N+2 technology. Concerns about yield remain, with some 910C units allegedly containing chips made by TSMC for the firm Sophgo. Huawei has denied using components from TSMC.

The US Commerce Department is currently examining the interactions between TSMC and Sophgo after discovering a Sophgo chip within Huawei's previous 910B processor. TSMC has consistently asserted that it has not supplied Huawei since 2020, remaining committed to export compliance.

By the end of 2023, Huawei started distributing initial samples of the 910C to selected tech firms and began accepting orders. The advisory firm Albright Stonebridge Group has posited that due to ongoing restrictions on US chip exports, the 910C could become the preferred choice for Chinese enterprises aiming to construct substantial AI models or enhance their computation capabilities.

Even though the Huawei Ascend 910C does not compete with Nvidia regarding power efficiency or cutting-edge technology, it represents a larger shift. Chinese firms are increasingly focused on developing indigenous alternatives to foreign technologies, even if such methods are less sophisticated yet provide comparable results.

As the international demand for AI grows and export limits tighten, Huawei's capacity to offer scalable AI hardware within the country may influence China's AI landscape by helping developers ensure stable supply lines and lessen geopolitical disruptions.

(Photo via Unsplash)