4月8日上午，由北京大学集成电路学院、集成电路学院高精尖创新中心、北京大学国家集成电路产教融合创新平台、集成电路科学与未来技术北京实验室、后摩尔时代微纳电子学科创新引智基地、北京大学校友会半导体分会联合主办的“未名·芯”论坛系列讲座第五十七期在北京大学微纳电子大厦103报告厅成功举办。本期由瑞士洛桑联邦理工学院的David Atienza Alonso教授带来题为“Accelerator-Centric Edge AI Architectures for Low-Power and Personalized Wearables”的精彩讲座，讲座由助理教授马宇飞主持。
在报告的开场部分，Atienza教授首先概述了边缘计算在现代物联网和可穿戴设备中的重要性，并强调了边缘AI架构在低功耗、个性化医疗监控等应用中的巨大潜力。他指出，现有的边缘端AI架构，如ASIC、PIM和CGRA等，虽然在能效方面已取得显著进展，但近年来其性能提升已达到相对瓶颈。为进一步增强芯片性能并降低能耗，需要结合应用领域专业知识，对人工智能系统进行软硬件协同优化。
在讲座中，Atienza教授分享了其团队在可穿戴设备中实现边缘AI的最新研究成果。他介绍了一种基于生物信号处理的事件驱动型边缘AI架构，能够根据不同生物信号（如心电图ECG、脑电图EEG等）的特性，动态调整计算资源，从而优化功耗和性能。此外，Atienza教授还分享了其团队将边缘AI与云端大数据和联邦学习相结合的相关工作与思考，旨在提升边缘端健康监测系统的性能并满足个性化需求。
此后，Atienza教授介绍了团队开发的全新开源边缘AI框架X-HEEP，并重点介绍了该框架在健康监测领域的应用——HEEPocrates系统。该系统在X-HEEP框架中集成了CGRA（Coarse-Grained Reconfigurable Accelerator）和存内计算。这种异构计算系统使用CGRA计算执行频繁执行的 DCFG（确定上下文无关文法，如循环体），使用存内计算模块计算常规AI算子，结合了CGRA的灵活性与存内计算的能效优势，提升了系统性能，展现了其在边缘AI应用中的巨大潜力。
最后，Atienza教授分享了他对医疗领域的新一代边缘AI架构的思考，认为可穿戴多设备多模态健康监测系统，以及实现高效设备端训练的AI架构，能够提高医疗健康监测的精度，是很有应用前景的研究方向。
提问环节，Atienza教授与现场的师生进行了深入的互动和讨论，回答了关于边缘AI架构设计、设备端学习、X-HEEP系统使用等方面的问题，启发了他们在未来科研工作中的思考。
个人简介：
David Atienza is a professor of Electrical and Computer Engineering, heads the Embedded Systems Laboratory (ESL), and is the Associate Vice President of Research Centers and Platforms for the period 2024-2028 at EPFL, Switzerland. His research interests include system-level design methodologies for multi-processor system-on-chip (MPSoC) targeting low-power Cyber-Physical Systems (CPS) and energy-efficient computing servers. His latest works include new 2.5D/3D power/thermal-aware design and architectures for MPSoCs targeting edge AI systems, as well as HW/SW co-design and AI-based multi-level optimization for sustainable computing in the Internet of Things (IoT) context.
Prof. David Atienza has co-authored over 450 papers, one book, and 14 patents in these previous areas. He has also received multiple recognitions and awards, among them the IEEE/ACM HW/SW Co-Design Conference (CODES-ISSS) 2024 Test-of-Time Award for the most influential paper in the last 15 years, the ICCAD 10-Year Retrospective Most Influential Paper Award in 2020, the Design Automation Conference (DAC) Under-40 Innovators Award in 2018, and IEEE CEDA and ACM SIGDA Early Career Awards on EDA tools and systems research. He is a Fellow of IEEE, Fellow of ACM, and was the Chair of the European Design Automation Association (EDAA) since 2022 until 2024. He is currently the Editor-in-Chief of IEEE Trans. on CAD (T-CAD) and ACM Computing Surveys.
