Optimized Distance Calculation Support for HBM PIM(Processing-In-Memory)
Nahyeon Kim, Sujin Kim, Haechannuri Noh, Min Jung, Huijin Roh, Ji-Hoon Kim · 2025
In this paper, we propose an extension of the Processing-In-Memory (PIM) architecture's instruction set to optimize distance calculations. The existing PIM architecture, while efficient for neural network tasks, exhibits limitations in handling distance calculations due to frequent data transfers between memory banks. To address this, we introduced the custom instruction, reducing intermediate data storage needs and improving computation efficiency. Simulations using random matrices showed that our approach decreases cycle count by up to 44 %, with more significant performance gains for larger dataset. This extension demonstrates the potential of enhanced PIM architectures for efficient memory-bound operations.