Huawei plans to begin mass production and shipment of its latest AI chip, the Ascend 910C, next month. While it achieves high performance, it has issues with power efficiency compared to NVIDIA products due to manufacturing process constraints.
🟦 Huawei’s “Ascend 910C” to be shipped in mass production: A new AI chip with doubled performance through structural evolution
Huawei’s “Ascend 910C” is a high-performance GPU that achieves double the computing power and memory capacity through a structural design that integrates two existing 910B chips. It is designed to be suitable for both AI inference and training, and aims for performance comparable to NVIDIA’s H100.
Samples have already been provided to some companies, and orders have begun. However, since the technology used is TSMC’s 7nm process technology and the manufacturing node is not the latest, it has been pointed out that it is inferior to NVIDIA products in terms of power efficiency.
🟦 The rise of Chinese companies brought about by US export restrictions
The background to this movement is the strict export restrictions imposed by the US government. NVIDIA’s H100, B200, and even the latest H20 chips have been blocked from the Chinese market, and the shortage of AI chips in the country is becoming serious.
To fill this void, Chinese companies such as Huawei, Moore Threads, and Iluvatar CoreX are accelerating GPU development. The 910C is one of them, and it is rapidly gaining presence in China as an option for AI inference processing. However, Huawei’s chips are said to contain parts made by TSMC, and about 3 million chips procured through the intermediary Sophgo are the subject of an investigation by the US Department of Commerce. TSMC claims that it has not shipped to Huawei since 2020, but the actual supply situation is unclear.
🟦 Summary
Huawei’s “Ascend 910C” is a Chinese-made AI chip that has achieved NVIDIA-class performance through ingenuity in structural design. Full-scale shipments began in May, and it can be said to be a symbol of China’s technological independence under sanctions. On the other hand, issues remain, such as reduced power efficiency due to the absence of advanced nodes and continued dependence on TSMC.
I was reminded that China has many issues that need to be resolved, including artificial intelligence, which requires large amounts of computing resources. The attitude of prioritizing computing performance over power efficiency may be a manifestation of the national will to “solve what needs to be solved now.”
