MS Project Defense -- Andrew Huang

Date: 
Friday, August 31, 2018 - 11:00am
Location: 
HFH 1132
Title: 
Applying Approximate Adders to AV1 for Energy-Efficient Video Compression
Committee: 
Yufei Ding (Chair), Ling-Qi Yan

Abstract:

 

Video applications have become proliferated on the Internet over the last decade, driving the demand for high efficiency video compression technology. AV1 is a new and emerging video codec that has superior compression rates over current video standards. Nevertheless, it has not been adopted into mainstream applications due to its high energy costs. To improve the energy efficiency of AV1, we introduce a novel idea to utilize approximate computing on AV1’s encoding process.

 

Through detailed analysis, we determine where to apply approximations by finding out which part in the encoding process is the most costly and error resilient. Then, we apply a low-power approximate addition method to greatly improve the energy-efficiency, without inducing significant error on the compressed video quality. Confirming our idea, our experiments show that our method could reduce the energy cost of AV1 by 2.5x while still maintaining 95% accuracy. Additionally, we also compare our results with HEVC, a competing video codec, and observe that AV1 could achieve higher accuracy while consuming equivalent energy.

 

Everyone Welcome!

Everyone welcome!