Presentation
ADROIT: An Adaptive Dynamic Refresh Optimization Framework for DRAM Energy Saving in DNN Training
Time
Location
Event Type
Research Manuscript
Hosted in Virtual Platform
AI/ML System Design
Design
DescriptionTo achieve high accuracy, DNN training usually consumes and generates myriads of data, which requires a large DRAM for efficient processing.
The refresh power consumption in large DRAM has become a severe problem.
We propose ADROIT, an adaptive dynamic refresh optimization framework for various DNNs and processing platforms.
ADROIT takes data idle time, lifetime and size into consideration, and dynamically adjusts the refresh rates for different types of data according to runtime loss feedback in DNN training.
Experimental results show that ADROIT reduces the refresh energy by up to 98.9% in DNN training, while maintaining the accuracy.
The refresh power consumption in large DRAM has become a severe problem.
We propose ADROIT, an adaptive dynamic refresh optimization framework for various DNNs and processing platforms.
ADROIT takes data idle time, lifetime and size into consideration, and dynamically adjusts the refresh rates for different types of data according to runtime loss feedback in DNN training.
Experimental results show that ADROIT reduces the refresh energy by up to 98.9% in DNN training, while maintaining the accuracy.