Memory-Centric AGI Computer Design: From Human Mind to Cognitive Microarchitectures
Abstract:
The pursuit of Artificial General Intelligence (AGI) requires cognitive microarchitectural processors that can overcome the severe energy and latency bottlenecks of current data-driven AI systems. This talk introduces a comprehensive, memory-centric hardware approach to AGI, structured around four foundational cognitive pillars: perceiving, understanding, deciding, and learning. To enable efficient perception, we explore Processing-in-Memory (PIM) innovations, including digital dynamic-logic SRAM and 3D BEOL cache architectures. For understanding complex relationships, we present hardware accelerators for graphical and probabilistic models, utilizing a Coarse-Grained Compressed Sparse Row (CG-CSR) format and dedicated Markov Chain Monte Carlo (MCMC) sampling. To address the slow and power-hungry nature of Large Language Model (LLM) reasoning, we introduce Cognitive Processing Units (CoPUs) that execute real-time, low-power symbolic decisions through parallel knowledge search and hardware Rete algorithms. Finally, we discuss accelerating Reinforcement Learning (RL) via the neuro-symbolic fusion of CoPUs with Neural Processing Units (NPUs) and FPGA-based environment accelerators like PEARL. By rethinking computer architecture from the ground up, we can build efficient, scalable cognitive microarchitectures that pave the way for next-generation AGI systems
Biodata:
Bonan Yan is an assistant professor at Institute for Artificial Intelligence, Peking University. He obtained his Ph.D. degree in Electrical and Computer Engineering from Duke University in 2020. His research fields include: artificial general intelligence processors and processing-in-memory (PIM) VLSI design and (mainly focusing on emerging technologies such as RRAM, PCM, MRAM, 3D DRAM and flexible IC design). His is the inventor of RRAM-based universal Ising machine and PIM-based Markov Chain Monte Carlo (MCMC) accelerator. He has published more than 60 papers in top-tier journals and conferences, including Nature, Nature Electronics, ISSCC, VLSI, DAC, HPCA, etc. He was awarded ACM SIGDA Outstanding New Faculty Award in 2024.