Presentation
PiMM-NoC: Process-in-Memristor-Memory NoC with RL Mapping Framework for Versatile AI Models
DescriptionProcess-in-memristor-memory (PiMM) offers promising in-/near-memory compute capability. However, existing PiMM designs support only a narrow set of model types and depend on coarse, static mapping strategies that overlook the vast design space, resulting in limited adaptability, scalability, and overall performance.
In this work, we propose a Process-in-Memristor-Memory Network on Chip (PiMM-NoC) architecture combined with a reinforcement learning (RL) based mapping framework to optimize on-chip latency for versatile AI workloads. PiMM-NoC integrates two types of tiles: PiMM tiles for weight-stationary operations and PU tiles for non-weight-stationary and nonlinear computations. A cycle-accurate architecture simulator is incorporated into an end-to-end hardware–software co-design framework that uses Monte Carlo Tree Search (MCTS) to automatically search efficient mapping strategies.
Experiments show that PiMM-NoC with the RL-based mapping framework achieves up to 3.45× speedup on DNNs and 3.85× on LLMs over existing mapping strategies, and up to 71.6× higher performance and 6.7× better energy efficiency compared to state-of-the-art AI accelerators.
In this work, we propose a Process-in-Memristor-Memory Network on Chip (PiMM-NoC) architecture combined with a reinforcement learning (RL) based mapping framework to optimize on-chip latency for versatile AI workloads. PiMM-NoC integrates two types of tiles: PiMM tiles for weight-stationary operations and PU tiles for non-weight-stationary and nonlinear computations. A cycle-accurate architecture simulator is incorporated into an end-to-end hardware–software co-design framework that uses Monte Carlo Tree Search (MCTS) to automatically search efficient mapping strategies.
Experiments show that PiMM-NoC with the RL-based mapping framework achieves up to 3.45× speedup on DNNs and 3.85× on LLMs over existing mapping strategies, and up to 71.6× higher performance and 6.7× better energy efficiency compared to state-of-the-art AI accelerators.
Event Type
Research Manuscript
TimeWednesday, July 2911:15am - 11:30am PDT
LocationMtg Room 203AB