1. 주제: Beyond Intelligent Computing: Scalable, Efficient, and Trustworthy AI Accelerators

2. 연사: 김보경 교수, Rutgers University

3. 장소: J107호

4. 강의 개요

Next-generation AI accelerators must deliver not only raw speed, but also scalability and efficiency across diverse workloads. This talk begins with our latest research on novel architectures and optimized dataflows (i.e., computation–communication co-optimization) that break traditional performance–energy trade-offs. To demonstrate the real-world impact of these technologies, I will introduce our custom seizure classification chip, capable of real-time EEG analysis and achieving high accuracy under tight power constraints for clinical use. Building on this foundation, we turn to the emerging imperative of trustworthiness: as AI and ML systems increasingly underpin critical decision-making, ensuring reliability, robustness, and privacy becomes paramount, alongside efficiency. I will present an accelerator architecture tailored for differentially private training, enabling privacy-preserving machine learning without compromising accuracy. Together, these advances illustrate how scalable, efficient, and trustworthy AI accelerators can move beyond intelligent computing toward safe and impactful deployment in life-critical domains.