Presentation
A Convolution Neural Network Accelerator Design with Weight Mapping and Pipeline Optimization
DescriptionIn-memory computing (IMC) convolution neural network (CNN) accelerators overcome von Neumann's bottleneck by in-situ computing. While the improvement of energy efficiency and throughput suffers from challenges from the access and movement of input and output data. A CNN accelerator design with overall system integration insight is desired. In this work, we propose a CNN accelerator design named MDopt that fully considers buffer access energy and latency. It includes a novel CNN mapping method and a pipeline optimization scheme. Experiments show that MDopt can reduce 49% and 22% energy consumption and improve 7.0х and 12.6х throughput for ResNet18 and VGG19, respectively.
Event Type
Research Manuscript
TimeTuesday, July 11th2:10pm - 2:25pm PDT
Location3010, 3rd Floor
Design
In-memory and Near-memory Computing Architectures, Applications and Systems