Close

Presentation

EMS: Efficient Memory Subsystem Synthesis for Spatial Accelerators
TimeTuesday, July 12th11:37am - 12pm PDT
Location3004, Level 3
Event Type
Research Manuscript
Keywords
RTL/Logic Level and High-level Synthesis
Topics
AI
EDA
DescriptionWe propose EMS, an efficient memory subsystem automatic generator for spatial accelerators. We use space-time transformation (STT) to analyze both PE-level and memory-level data reuse. Based on STT, we develop an algorithm to automatically generate the layout of the multi-banked scratchpad memory based on the memory-level reuse, and also generates the data mapping and access controller for the memory. The memory subsystem supports multiple PE-memory interconnection topologies including direct, multicast and rotated connection. Experiment shows that our memory generator requires 22% to 78% smaller memory size than previous works for the same tensor algebra workload, and achieves comparable hardware performance.