Huawei is set to revolutionize the technological landscape by launching mass shipments of its Ascend 910C. In light of the repercussions of U.S. restrictions, this initiative marks a significant step in the quest for autonomy of Chinese companies in the field of semiconductors. With a local alternative to imported technology, Huawei is making significant strides in artificial intelligence systems, generating considerable repercussions in the global market.
Start of mass shipments of the Ascend 910C
Huawei is preparing to proceed with mass shipments of its new AI chip Ascend 910C, an initiative scheduled for next month, according to well-informed sources. This advancement is a decisive element for Chinese companies, which are seeking local alternatives to American semiconductors.
Background of U.S. restrictions
Huawei’s decision echoes a tense situation, marked by increasing American restrictions on access to Nvidia hardware. Recently, the U.S. government notified Nvidia of the need for an export license to sell its AI chip H20 to China. Thus, Chinese developers find themselves searching for solutions capable of supporting large-scale training and inference workloads.
Features of the Ascend 910C
Although this new chip is not based on the most advanced manufacturing processes, it offers a viable alternative. The 910C presents itself as a double-package version of the previous model, the 910B, thus providing doubled performance and memory capacity. Experts report that its yield compares to that of the Nvidia H100.
Technical approach adopted by Huawei
Rather than relying on cutting-edge manufacturing processes, Huawei is taking a pragmatic approach. This strategy involves integrating multiple chips with high-speed optical interconnections to enhance performance. This method is at the heart of the CloudMatrix 384 system, an AI platform at the rack scale designed for training complex models.
Architecture and specifications of the CloudMatrix 384
The CloudMatrix 384 system includes 384 Ascend 910C chips, organized across 16 racks made up of 12 compute racks and four network racks. Unlike copper-based infrastructures, Huawei’s platform uses optical interconnections to allow high-bandwidth communication between different components. SemiAnalysis analyses indicate that the architecture incorporates 6,912 800G LPO optical transceivers, forming a meshed optical network.
Comparative performances
This innovation allows Huawei’s infrastructure to reach about 300 petaFLOPs of BF16 computing power, thus surpassing Nvidia’s GB200 NVL72 system, which achieves around 180 BF16 petaFLOPs. The CloudMatrix also stands out with higher memory bandwidth and capacity, boasting more than double the bandwidth and over 3.6 times the high bandwidth memory (HBM) capacity.
Comparative efficiencies and challenges
The performance advantages come with drawbacks. Huawei’s system is estimated to be 2.3 times less efficient per floating-point operation than Nvidia’s GB200 and exhibits lower energy efficiency per unit of memory bandwidth. Despite this performance ratio, Huawei’s infrastructure still provides the necessary elements for training advanced AI models at scale.
Production and components
The manufacturing of some essential components for the 910C is attributed to China’s largest foundry, SMIC, using its 7nm N+2 process. However, yield concerns remain, and some 910C units may include chips produced by TSMC for the Chinese company Sophgo. Huawei has insisted that it does not use components made by TSMC.
Investigation and regulatory implications
The U.S. Department of Commerce is currently investigating the links between TSMC and Sophgo, after a chip designed by Sophgo was found in Huawei’s previous processor, the 910B. TSMC has claimed that it has not supplied Huawei since 2020 and continues to comply with export regulations.
Impact on the technology industry in China
During the year 2023, Huawei began distributing early samples of the 910C to selected technology companies while opening its order books. The consulting firm Albright Stonebridge Group believes that this chip could become the go-to choice for Chinese companies engaged in building large-scale AI models or deploying inference capabilities, given the ongoing export controls on American chips.
Toward technological autonomy
The launch of the Ascend 910C, while not competing with Nvidia in terms of energy efficiency or process technology, marks a broader trend. Chinese technology companies are accelerating the development of domestic alternatives to foreign components, even if this involves using less advanced methods to achieve similar results.
As global demand for AI explodes and export restrictions tighten, Huawei’s ability to provide a scalable AI hardware solution on Chinese soil could shape the future of artificial intelligence in the country, particularly for developers seeking to secure long-term supply chains and reduce their exposure to geopolitical risks.
For further information and analysis on this technology, the AI & Big Data Expo offers rich sessions in Amsterdam, California, and London, integrating other major events such as the Intelligent Automation Conference, BlockX, and Cyber Security & Cloud Expo.
FAQ about Huawei’s shipment of the Ascend 910C
What is Huawei’s Ascend 910C chip?
The Ascend 910C chip is a processing unit dedicated to AI, designed to deliver high performance in computing, with a dual-package version that enhances power and memory capacity compared to its predecessor, the 910B.
When does Huawei plan to start mass shipments of the Ascend 910C?
Huawei is expected to start its mass deliveries of the Ascend 910C chip next month, following previous limited shipments.
Why is Huawei launching the Ascend 910C despite U.S. restrictions?
Huawei is responding to the growing need for domestic solutions in China due to restrictions on access to American components, particularly those from Nvidia.
What advantages does the Ascend 910C chip present compared to Nvidia solutions?
The Ascend 910C chip offers computing performance of 300 petaFLOPs, surpassing Nvidia’s GB200 NVL72 system, focusing on an architecture that uses optical interconnections for superior bandwidth.
What technologies are used in Huawei’s CloudMatrix 384 system?
The CloudMatrix 384 system features 384 Ascend 910C chips with optical interconnections and is configured into 16 racks to provide optimized processing capabilities for large-scale AI models.
What challenges relate to the efficiency of the Ascend 910C chip?
While performing well, the efficiency per floating-point operation of the 910C is lower, achieving 2.3 times less efficiency compared to some Nvidia products, raising concerns about energy consumption.
How does the current situation affect the Chinese chip market?
Export restrictions on American chips are driving Chinese companies to develop local alternatives such as the Ascend 910C, thereby enhancing the country’s technological autonomy.
What is the impact of SMIC’s production on the Ascend 910C?
SMIC, the largest chip manufacturer in China, produces some of the components of the 910C using its 7nm N+2 process, although concerns about yield reliability persist.
How is the U.S. government reacting to Huawei’s use of TSMC technologies?
The U.S. Commerce Department is investigating the relationships between TSMC and Sophgo after chips designed by Sophgo were found in Huawei’s previous processors, which could have ramifications for Huawei’s supply chain.
