Huawei unveils Ascend 910D AI chip with 15% performance improvement, leveraging SMIC’s 7nm process despite US sanctions, as China expands semiconductor funding and EU accelerates chip subsidies.
On 12 June 2024, Huawei launched its Ascend 910D AI processor boasting 640 TOPS compute power, manufactured through SMIC’s 7nm node with 60% yield rates reported by Reuters on 10 June.
Geopolitical Context of Semiconductor Development
Huawei’s Ascend 910D emerges as China responds to expanded US Entity List restrictions from 11 June targeting 37 AI-related entities. The $47 billion semiconductor fund expansion, announced last week by China’s State Council, directly supports domestic production capabilities. ‘This represents textbook semiconductor nationalism,’ notes TechInsights analyst Dan Hutcheson. ‘China is prioritizing functional chips over perfect yields in strategic sectors.’
Technical Specifications and Production Challenges
The 910D’s architecture shows similarities to Nvidia’s A100 in tensor core configuration, though modified for SMIC’s 7nm process. While achieving 640 TOPS (15% over previous gen), SMIC’s yield rates remain at 60% compared to TSMC’s 95% for equivalent nodes. ASML confirmed on 14 June it halted EUV maintenance contracts for Chinese fabs under updated Dutch export rules.
Global Market Implications
EU Commissioner Thierry Breton announced €8.1B in neuromorphic computing subsidies on 13 June, stating: ‘We cannot rely on single-region supply chains.’ Meanwhile, Alibaba Cloud has committed to deploying 50,000 Ascend 910D units by Q3 2024 for large language model training.
The Ascend 910D development follows China’s pattern of iterative improvements under constraints – SMIC achieved 7nm production in 2022 using deep ultraviolet (DUV) lithography after being denied extreme ultraviolet (EUV) tools. This mirrors previous adaptations like Huawei’s HarmonyOS replacement for Android after 2019 trade restrictions.
Historical data shows China’s semiconductor imports dropped 18% YoY in Q1 2024, while domestic production capacity reached 23% of total consumption, up from 15% in 2022. However, the 910D’s 300W power consumption remains 40% higher than Nvidia’s H100 for comparable AI workloads, highlighting persistent efficiency gaps.
