Huawei is poised to initiate large-scale shipments of its Ascend 910C AI chip as early as next month, according to insiders. While initial deliveries have already occurred in limited quantities, mass deployment signifies a crucial advancement for Chinese companies in their quest for domestic alternatives to U.S.-made semiconductors.
This development arises amid increasing restrictions on Chinese developers’ access to Nvidia hardware. Recently, the U.S. government mandated that sales of Nvidia’s H20 AI chip to China require an export license, prompting developers to seek viable options for large-scale training and inference tasks.
The Ascend 910C chip, while not fabricated using the latest process technology, offers a significant workaround. Essentially a dual-package version of the earlier 910B, it features two processors that double its performance and memory capacity.
Reports suggest its performance rivals that of Nvidia’s H100. Huawei’s strategy leans on a brute-force design, integrating multiple chips and high-speed optical interconnects to enhance performance.
This is fundamental to Huawei’s CloudMatrix 384 system, which forms a full rack-scale AI platform aimed at training large models. The CloudMatrix 384 incorporates 384 Ascend 910C chips across 16 racks, comprising 12 for computing and four for networking.
In contrast to conventional copper-based systems, Huawei’s designs utilize optical interconnects, providing high-bandwidth communication between components. Analysis indicates that the architecture employs 6,912 optical transceivers to create a high-capacity optical network.
While the CloudMatrix achieves around 300 petaFLOPs of BF16 computing power—exceeding Nvidia’s GB200 NVL72 system, which delivers about 180 BF16 petaFLOPs—it faces efficiency challenges. Predictions indicate that Huawei’s system is 2.3 times less efficient per floating-point operation than Nvidia’s offering and struggles with power efficiency for memory bandwidth and capacity.
Nonetheless, it still supports the infrastructure necessary for training sophisticated AI models on a large scale. China’s largest chipmaker, SMIC, is producing key components of the 910C utilizing a 7nm N+2 process, although yield levels are under scrutiny.
Some models reportedly include chips from TSMC, produced for the Chinese company Sophgo, although Huawei has denied any reliance on TSMC materials. The U.S. Commerce Department is examining the relationship between TSMC and Sophgo following the discovery of a Sophgo-designed chip in Huawei’s previous processor model.
In late 2023, Huawei dispatched early samples of the 910C to select tech firms and opened up for orders. Consulting experts believe this chip could become a primary option for Chinese enterprises engaged in developing large AI models or extending inference capacities, especially in light of ongoing export restrictions on U.S. semiconductor products.
Although the Ascend 910C may not match Nvidia in terms of power efficiency or process technology, it represents a significant shift towards homegrown technological solutions. As global AI demand escalates and export controls tighten, Huawei’s ability to provide scalable AI hardware could considerably influence China’s AI sector, aiding developers in establishing resilient supply chains while minimizing geopolitical risks.