Hybrid Talk + Discussion Prof. Can Li: “Fault-Tolerant Analog Computing with Emerging Devices for Efficient, Trustworthy, and Scalable AI“

Virtual participation:
Join the talk in Zoom

We cordially invite you to the PGI-14 hosted hybrid talk by Prof. Can Li (University of Hong Kong, Department of Electrical and Electronic Engineering),

Abstract: The exponential growth of generative AI has placed unprecedented demands on computing hardware, pushing data center energy consumption toward terawatt scales. To address this, there is an urgent need to move beyond traditional von Neumann architectures. This talk explores brain-inspired, in-memory computing systems that leverage the physical properties of emerging devices, such as oxide memristors and 2D material-based transistors, to achieve high throughput and energy efficiency.

First, I will discuss the challenge of intrinsic imperfections in analog computing. While memristive crossbar arrays enable efficient parallel vector-matrix multiplication via Ohm’s and Kirchhoff’s laws, device variability and yield remain critical hurdles. I will present our recent work on Fault-Free Matrix Representation. Unlike traditional redundancy methods, this approach mathematically decomposes a target matrix into the product of two tunable sub-matrices. This adaptive representation bypasses stuck-at faults and eliminates the need for differential pairs, effectively doubling storage density. We demonstrate that this method achieves superior tolerance to high fault rates and significantly reduces the bit error rate in 6G wireless communication prototypes.

Second, I will introduce our efforts in building fully integrated spiking neural networks (SNNs). We have developed a system combining 180nm CMOS with TaOx memristor synapses and custom analog Spike Response Model (SRM) neurons. A key feature is the proportional time-scaling property of our analog neurons, which allows the processing of event-based data accelerated to microsecond scales using compact on-chip capacitors. This system achieves 93.06% accuracy on the DVS128 Gesture dataset with high energy efficiency. By bridging device physics with event-driven processing, we demonstrate a viable, scalable pathway for ultra-fast edge intelligence.

Finally, I will present advancements in 2D material-based electronics for in-memory searching and AI acceleration. We utilize atomically thin MoS2 flash devices with semimetal antimony (Sb) contacts, achieving record-high readout currents and sub-0.1 fJ energy consumption per search. Leveraging this platform, we developed a hardware-software co-design for Soft Decision Trees (SDT). By utilizing the inherent soft boundaries of analog devices, we achieve exceptional robustness against adversarial attacks, validating 96% accuracy on medical datasets. I will conclude with a Monolithic 3D (M3D) architecture for Transformer acceleration. By vertically integrating static RRAM weights with high-speed MoS2 eDRAM for dynamic attention Key-Value caching, we effectively overcome the bandwidth bottlenecks inherent to Large Language Model (LLM) inference.

About the speaker: Can Li’s research aims to explore and build the next-generation computing hardware based on post-CMOS emerging devices, e.g. memristors. Before joining HKU, Dr. Li worked at Hewlett Packard Labs in California, USA. He obtained his bachelor’s degree from Peking University and his Ph.D. from the University of Massachusetts Amherst. In recent years, he has published 82 peer-reviewed publications, including 19 research papers in Nature series journals. Based on his research achievements, after joining HKU, he has received the National Natural Science Foundation of China Excellent Young Scientist Fund, the Hong Kong Research Grants Council Early Career Award, and the Croucher Tak Wah Mak Innovation Awards.

We hope to see many of you!