The Ascend 910C represents Huawei's most advanced AI chip, built on SMIC's second-generation 7nm (N+2) process technology12. This powerful processor delivers approximately 800 TFLOPS of FP16 performance3 and achieves about 60% of Nvidia's H100 inference performance in DeepSeek model-based tasks12. The chip features a chiplet packaging design with approximately 53 billion transistors41 and utilizes HBM2e memory with around 3.2 TB/s memory bandwidth5.

Key specifications include:

• Architecture: Based on Huawei's proprietary Da Vinci architecture6

• Performance: 800 TFLOPS at FP163, with 320 TFLOPS of FP16 and 64 TFLOPS of INT8 performance for various AI tasks6

• Power consumption: Approximately 310 watts67

• Memory: High-bandwidth HBM2e memory6

• Software compatibility: Supports Huawei's MindSpore AI framework, TensorFlow, and PyTorch61

• Manufacturing: SMIC 7nm process, shifting from TSMC's N7+ used in previous generations12

Huawei's Ascend 910C represents a significant architectural evolution in China's AI chip landscape, achieving performance comparable to Nvidia's H100 by combining two 910B processors through advanced integration techniques.12 This design effectively doubles the computing power and memory capacity while enhancing support for diverse AI workloads. Beyond Huawei, China's domestic AI chip ecosystem includes several key players working to fill the gap created by U.S. export restrictions, including Baidu's Kunlunxin series, Alibaba's T-Head, and startups like Biren Technology and Horizon Robotics.3

The next-generation Ascend 920, built on SMIC's 6nm process (an upgrade from the 910C's 7nm), promises even greater capabilities with 900 TFLOPs of compute performance and 4TB/sec of memory bandwidth.45 Chinese companies are increasingly adopting these domestic alternatives, with some reporting up to 20% cost reductions compared to Nvidia-based systems while maintaining comparable performance.6 This shift toward technological self-reliance has accelerated as Chinese authorities informally advise local companies to prioritize domestic chips over foreign alternatives, despite the engineering challenges involved in transitioning away from Nvidia's ecosystem.7

The U.S. government recently imposed new licensing requirements for AI chip exports, particularly affecting Nvidia's H20 chips destined for China. According to an SEC filing, Nvidia must now secure licenses to export H20 integrated circuits to China, Hong Kong, Macau, and other arms-embargoed nations, with the requirement remaining "in effect indefinitely."1 This restriction aims to prevent these chips from being "utilized or redirected to a supercomputer in China."1

The financial impact is substantial, with Nvidia projecting approximately $5.5 billion in charges related to "inventory, purchase obligations, and associated reserves" for the affected chips.1 These controls are part of broader U.S. efforts to maintain technological leadership in AI development and chip design.2 The restrictions create an opportunity for Huawei's 910C chips to fill the market gap in China, as Chinese companies face increasing difficulty accessing advanced foreign AI chips that could compete with Huawei's offerings.3