Table of Contents
Key Takeaways
- Huawei’s Ascend 910C delivers up to 800 TFLOPS FP16 performance, rivaling Nvidia’s H100 in key AI workloads.
- U.S. export controls on Nvidia’s H20 chips have created a market gap that Huawei and other Chinese chipmakers are rapidly filling.
- The 910C leverages advanced chiplet packaging and SMIC’s 7nm process, highlighting China’s progress in semiconductor manufacturing.
- Major Chinese tech firms are transitioning to domestic AI chips, reporting cost savings and improved supply chain security.
- The upcoming Ascend 920 chip on a 6nm process signals Huawei’s commitment to ongoing innovation in AI hardware.
- Government policies and industry collaboration are accelerating China’s shift away from foreign AI chip dependence.
- Domestic alternatives like Baidu’s Kunlunxin and Alibaba’s T-Head are also gaining traction in specialized AI markets.
- The 910C’s compatibility with popular AI frameworks ensures smooth integration for developers and enterprises.
Huawei Ascend 910C: Technical Innovation and Performance
- The Ascend 910C is Huawei’s most advanced AI processor to date, built using SMIC’s second-generation 7nm (N+2) process technology. This marks a significant milestone in China’s ability to produce high-performance chips domestically.
- At the heart of the 910C is a chiplet-based design that integrates two Ascend 910B processors, effectively doubling both computing power and memory bandwidth compared to its predecessor. This innovative approach enables the 910C to achieve up to 800 TFLOPS of FP16 performance and 3.2 TB/s memory bandwidth, putting it in direct competition with Nvidia’s H100.
- The chip features approximately 53 billion transistors and utilizes HBM2e high-bandwidth memory, supporting the massive data throughput required for large language models and deep learning workloads.
- With a power consumption of around 310 watts, the 910C is also more energy-efficient than many competing solutions, making it attractive for large-scale data centers and AI clusters.
- Compatibility with leading AI frameworks—including Huawei’s MindSpore, TensorFlow, and PyTorch—means that organizations can transition their existing workloads with minimal friction.
U.S. Export Restrictions: A Catalyst for Domestic Innovation
- In response to growing concerns about China’s access to advanced AI technology, the U.S. government has imposed strict export controls on high-performance chips, most notably Nvidia’s H20. These restrictions require U.S. companies to secure licenses for exports to China, Hong Kong, and Macau, with the rules remaining in effect indefinitely.
- Nvidia has projected a $5.5 billion financial impact due to inventory and purchase obligations for chips affected by these new regulations. The intent behind these controls is to prevent advanced AI chips from being used in Chinese supercomputers or military applications.
- This regulatory environment has created a significant opportunity for Chinese chipmakers like Huawei, who are now able to fill the market gap with competitive domestic alternatives.
- As a result, major Chinese tech firms—including ByteDance, Tencent, and Alibaba—are accelerating their adoption of homegrown AI hardware, reducing their reliance on foreign suppliers and mitigating supply chain risks.
China’s Expanding Domestic AI Chip Ecosystem
- Huawei is not alone in the race to develop advanced AI chips. Companies such as Baidu (with its Kunlunxin series), Alibaba (T-Head), Biren Technology, and Horizon Robotics are all contributing to a rapidly evolving domestic ecosystem.
- These firms are targeting a variety of AI applications, from cloud computing and large-scale language models to edge AI and autonomous vehicles.
- Chinese authorities have informally advised local companies to prioritize domestic chips over foreign alternatives, further accelerating the shift toward technological self-reliance.
- Reports from early adopters indicate that switching to domestic AI chips can reduce costs by up to 20% compared to Nvidia-based systems, while maintaining comparable performance for many workloads.
- The competitive landscape is fostering innovation, with each company focusing on different market segments and technical strengths.
The Road Ahead: Ascend 920 and Beyond
- Huawei’s roadmap extends well beyond the 910C. The next-generation Ascend 920 chip, currently in development, will be built on SMIC’s 6nm process and is expected to deliver 900 TFLOPS of FP16 compute performance and 4 TB/s of memory bandwidth.
- Early benchmarks suggest the Ascend 920 could be 30-40% faster than the 910C, making it a formidable competitor in the global AI chip market.
- Huawei is also investing heavily in software, with its MindSpore framework and CloudMatrix platform designed to maximize the performance and scalability of its hardware.
- The company’s partnerships with leading Chinese tech firms and cloud providers are ensuring that its chips are widely adopted and supported across the country’s digital infrastructure.
Huawei’s Ascend 910C is a landmark achievement in China’s quest for AI chip independence, offering performance and efficiency that rival the world’s best. As U.S. export controls tighten, China’s domestic semiconductor industry is poised for rapid growth and innovation, reshaping the global AI landscape.
