For further information, please contact the General chair: Pierre-Emmanuel Gaillardon
Heterogeneous Processing-in-memory for Energy-efficient CNN Training
Convolutional neural networks (CNNs) have been adopted in a wide range of application domains, such as image classification, speech recognition, object detection, and computer vision. CNNs can be trained in various computer systems. However, the overhead of data movement between processors and the memory is becoming one of the most critical bottlenecks, as the training models scale up. To address the issue, we proposes a hardware/software coordinated design to accelerate CNN training by processing-in-memory (PIM). Our PIM architecture design adopts hundreds of fix-function arithmetic units and four ARM-based programmable cores integrated on the logic layer of 3D die-stacked memory. Furthermore, we develop a set of software schemes to determine and offload the proper CNN training operations that directly execute on the memory side in either fixed-function PIMs, programmable PIMs, or a combination of the two. Our PIM-based system design can accelerate native training models written on TensorFlow.
Jishen Zhao is an Assistant Professor in the Computer Science and Engineering Department at University of California, San Diego. Her research falls primarily in the area of computer system and architecture, with an emphasis on memory and storage systems, high-performance computing, and energy efficiency. She is also interested in electronics design automation and VLSI design for three-dimensional integrated circuits and nonvolatile memories. Before joining UCSD, she was an Assistant Professor at UC Santa Cruz, and she worked at HP Labs before joining UCSC.