Semiconductor Behind China's "DeepSeek Shock," the "Ascend 910C": Clear Technological Limitations [Tech Talk]

Structure Connecting Two Previous-Generation Chips
Low Electricity Prices Offset Limitations

Ascend 910C graphics processing unit by Huawei (Chinese name: Sheng Teng·昇騰). Huawei

The recent emergence of the Chinese AI model "DeepSeek," which has shaken up the artificial intelligence (AI) industry, has also raised concerns in the hardware sector. This is because it was trained and inferred using the Chinese-made graphics processing unit (GPU), the Huawei Ascend (Chinese name: Sheng Teng·昇騰) 910C. The Ascend 910C is the third generation in Huawei's Ascend GPU series and is considered a cornerstone of technological self-sufficiency in China, where direct imports of high-performance NVIDIA GPUs are not possible. Recently, Western semiconductor experts have highlighted the structure and limitations of the Ascend 910C.

Chiplet Technology at the Core of Next-Generation Chinese GPUs

Chip structure of Ascend 910C. It is a form where two Ascend 910B chips from the previous generation are connected side by side as chiplets. Photo by X Capture

The Center for Strategic and International Studies, a U.S. national strategy think tank, obtained and published detailed specifications of the Huawei 910C this past March. Since then, Western tech experts have thoroughly analyzed the chip's structure and manufacturing methods.

According to reports from U.S. technology outlets such as Tom’s Hardware and SemiAnalysis, the Ascend 910C is essentially a large chip formed by connecting two previous-generation 910B chips together. This is possible due to the use of chiplet technology, which allows multiple semiconductor dies to function as a single chip. As a result, the 910C can theoretically deliver twice the performance of its predecessor without any significant process innovation.

The reason Huawei adopted this approach is likely due to technological limitations. The 910C, like the 910B, uses a 7-nanometer (nm) process. However, what is significant is that it uses the 7nm process of China’s state-owned semiconductor foundry SMIC, rather than Taiwan's TSMC. With this, the Ascend 910C achieves technological self-sufficiency by utilizing Chinese intellectual property (IP), Chinese semiconductor process equipment, and Chinese contract manufacturing.

Clear Limitations Due to Lack of Access to Advanced Global Technology

CloudMatrix 384, a cloud solution based on Ascend 910C. Although it occupies more space and has significantly lower watt-per-performance compared to NVIDIA products, it remains the only option for Chinese companies. Huawei

However, there are clear limitations to the approach taken with the Ascend 910C. The first is increased area. Since two chips are combined into one, the physical space occupied by the chip also increases. AI semiconductors must be installed by the tens of thousands in data centers, so an increase in chip die area means that server racks and other auxiliary equipment must also accommodate more space. This directly increases the burden on data center construction companies.

The second limitation is memory bandwidth. Unlike NVIDIA, Chinese companies do not have access to high-performance high-bandwidth memory (HBM) made in Korea, which creates a bottleneck. The Ascend 910C uses HBM2e. Its memory bandwidth per pin is 3.6Gbps, which is nearly half that of HBM4 (6.4Gbps) to be used in NVIDIA's next-generation chips. If the bandwidth, which is the standard for AI memory transfer speed, is low, AI training performance cannot be fully realized.

To overcome these limitations, Huawei employs expensive communication technology. On April 16, Huawei unveiled an AI cloud system called CloudMatrix 384. This system, equipped with thousands of Ascend 910C units, connects server racks exclusively with optical fiber instead of copper cables. Optical fiber transmits data faster than copper, but it is much more expensive, and as the communication distance increases, so does power consumption.

Taking all these various penalties into account, the power efficiency of the Ascend 910C is estimated to be about 60% of the computing performance per watt compared to the NVIDIA H100 GPU.

Low Electricity Prices Offset Inefficiency

View of Wujing Coal-Fired Power Plant in Shanghai, China. Photo by Yonhap News

Computing performance per watt is crucial for AI training and inference competitiveness. As AI continues to grow exponentially, only by suppressing development and operational costs through ultra-large data centers can companies remain competitive. If Chinese companies can achieve only 60% of the H100’s computing performance with the same 1 watt, they must compensate for the remaining 40% through additional data center investments. This also increases secondary costs, such as heat management, cooling water consumption, and the need for more operational staff, which could hinder Chinese AI companies.

However, for now, Chinese companies appear able to tolerate the inefficiency of CloudMatrix. Ironically, this is due to the sluggish Chinese economy.

According to Bloomberg in January of this year, China’s wholesale electricity price (per megawatt-hour, MWh) currently stands at $56 (about 80,000 won), significantly down from the 2022 average of $90.7 (about 128,000 won). This is because local governments have cut industrial electricity rates to prevent economic stagnation. Originally intended to support energy-intensive manufacturers, this policy has also provided windfall benefits to IT companies operating data centers.