PIM-FW: Hardware-Software Co-Design of All-pairs Shortest Paths in DRAM
Published in GLSVLSI '26, 2026
Overview
PIM-FW is a DRAM-based Processing-in-Memory framework for accelerating All-Pairs Shortest Paths using blocked Floyd-Warshall on HBM3. Through hardware-software co-design, it reduces data movement by performing memory-intensive computation near the data.
The hybrid PIM/PNM architecture uses in-bank Bank PEs for massively parallel min-plus updates and channel-level PEs for reduction and synchronization. Under the paper’s stated assumptions, it achieves an 18.7x speedup over an NVIDIA A100 baseline and up to 3200x lower modeled PIM-stack energy.
