Error Aware Computing
November 17, 2014
Operating logic at low, sub threshold voltages allows for interesting trade-offs between accuracy and energy consumption. While many signal and image processing applications can tolerate some errors, error-aware systems must be carefully architected to ensure that they deliver the expected energy savings. This talk describes work with Se Hun Kim and Saibal Mukhopodhyay on error-aware design of JPEG image compression systems. We describe an improved error model for sub threshold logic. Previous methods, based on a combinational model, overestimated errors. Our sequential model provides more accurate results. We developed models for sub threshold operation of JPEG compression. We showed that errors in the discrete cosine transform (DCT) result in increased file sizes from Huffman compression; the larger amounts of data that must be written to memory negate the energy savings from sub threshold operation of DCT. To solve this problem, we developed a new approach based on quantization. We also developed a memory architecture adapted to variable-width data.
Marilyn Wolf is Farmer Distinguished Chair and Georgia Research Alliance Eminient Scholar at the Georgia Institute of Technology. She received her BS, MS, and PhD in electrical engineering from Stanford University in 1980, 1981, and 1984, respectively. She was with AT&T Bell Laboratories from 1984 to 1989. She was on the faculty of Princeton University from 1989 to 2007. Her research interests include cyber-physical systems, embedded computing, embedded video and computer vision, and VLSI systems. She has received the ASEE Terman Award and IEEE Circuits and Systems Society Education Award. She is a Fellow of the IEEE and ACM and an IEEE Computer Society Golden Core member.