Presentation
DNNPhaser: Enhancing Data Locality Using Multiphase Ring Dataflow for Spatial Accelerators
DescriptionData rotation across spatial engines is a common strategy for reusing data from neighboring on-chip
engine buffers. However, the existing data rotation approach often results in inadequately sized
data tiles. In this paper, we introduce DNNPhaser, a framework designed to implement a multiphase
approach to 2D co-rotation. By preserving a phase difference between the rotation rings, DNNPhaser facilitates the coherent rotation of both input data rings and kernel data rings, ensuring the maintenance of appropriately sized data tiles. Experimental results demonstrate that DNNPhaser achieves a geometric mean EDP reduction of 24.8% for DNNs compared to Tangram on a spatial accelerator.
engine buffers. However, the existing data rotation approach often results in inadequately sized
data tiles. In this paper, we introduce DNNPhaser, a framework designed to implement a multiphase
approach to 2D co-rotation. By preserving a phase difference between the rotation rings, DNNPhaser facilitates the coherent rotation of both input data rings and kernel data rings, ensuring the maintenance of appropriately sized data tiles. Experimental results demonstrate that DNNPhaser achieves a geometric mean EDP reduction of 24.8% for DNNs compared to Tangram on a spatial accelerator.
Event Type
Work-in-Progress Poster
TimeWednesday, June 265:00pm - 6:00pm PDT
LocationLevel 2 Lobby
AI
Autonomous Systems
Cloud
Design
EDA
Embedded Systems
IP
Security