ICCAD 2026 Workshop: Memory-Centric Computing for LLM Inference

General Information

The rapid growth of large language models (LLMs) is creating unprecedented demands on memory capacity, bandwidth, energy efficiency, and scalability. As LLM inference scales toward long-context reasoning, high-throughput serving, and edge deployment, memory access and data movement have become major system bottlenecks, exposing the limitations of conventional GPU-centric architectures. Emerging paradigms such as processing-in-memory (PIM), near-memory and near-storage computing, 3D-stacked PIM, emerging non-volatile memory (NVM)-based computing architectures, heterogeneous computing, and hardware–software co-design offer promising solutions for scalable and energy-efficient LLM inference.

This workshop brings together researchers and practitioners across architecture, hardware design, EDA, systems, and AI to discuss the future of memory-centric computing for LLM inference. Topics include 2D/3D PIM, near-memory and near-storage computing, emerging NVM-based computing architectures, heterogeneous CPU/GPU/NPU–PIM systems, memory hierarchy, data movement and KV-cache optimization, compiler and runtime support, hardware–software co-design, and EDA and design-space exploration for memory-centric AI. By fostering interactions across architecture, hardware design, memory, systems, EDA, and AI communities, the workshop aims to identify key challenges, emerging opportunities, and future research directions for next-generation memory-centric AI computing infrastructures.