Event Date: June 10, 2021 16:15
Fighting Temperature: The Unseen Enemy for Neural processing units (NPUs)
Abstract - Neural processing units (NPUs) are becoming an integral part in all modern computing systems due to their substantial role in accelerating neural networks. In this talk, we will discuss the thermal challenges that NPUs bring, demonstrating how multiply-accumulate (MAC) arrays, which form the heart of any NPU, impose serious thermal bottlenecks to any on-chip systems due to their excessive power densities. We will also discuss how elevated temperatures severely degrade the reliability of on-chip memories, especially when it comes to emerging non-volatile memories, leading to bit errors in the neural network parameters (e.g., weights, activations, etc.). In this talk, we will also discuss: 1) the effectiveness of precision scaling and frequency scaling (FS) in temperature reductions for NPUs and 2) how advanced on-chip cooling using superlattice thin-film thermoelectric (TE) open doors for new tradeoffs between temperature, throughput, cooling cost, and inference accuracy in NPU chips.
Short bio - Dr. Hussam Amrouch is a Junior Professor at the University of Stuttgart heading the Chair of Semiconductor Test and Reliability (STAR) as well as a Research Group Leader at the Karlsruhe Institute of Technology (KIT), Germany. He earned in 06.2015 his Ph.D. degree in Computer Science (Dr.-Ing.) from KIT, Germany with distinction (summa cum laude). After which, he has founded and led the “Dependable Hardware” research group at KIT. Dr. Amrouch has published so far 115+ multidisciplinary publications, including 43 journals, covering several major research areas across the computing stack (semiconductor physics, circuit design, computer architecture, and computer-aided design). His key research interests are emerging nanotechnologies and machine learning for CAD. Dr. Amrouch currently serves as Associate Editor in Integration, the VLSI Journal as well as a guest and reviewer Editor in Frontiers in Neuroscience.
