High-Level Synthesis for Accelerator-Rich Computing
Speaker
Zhiru Zhang
School of ECE at Cornell University
Host
Professor Arvind
CSG - CSAIL - MIT
Abstract
Systems across the computing spectrum, from edge devices to cloud datacenters, are increasingly turning to specialized hardware accelerators for improved performance and energy efficiency. However, greater performance-per-watt comes at the cost of a much higher development effort. As the practice of traditional register-transfer-level (RTL) design has become unequivocally difficult, if not already unsustainable, high-level synthesis (HLS) has emerged as a promising approach to productive hardware specialization by enabling automatic generation of cycle-accurate RTL from untimed functional descriptions.
In this talk, I will review the progress we have made on HLS in recent years. Several real-life applications will be used as case studies to show the benefits and limitations of the current HLS tools. I will then mainly focus my talk on scheduling, which forms the algorithmic core to HLS. I will first describe a family of novel static scheduling algorithms devised for achieving high quality-of-results out-of-the-box. In particular, I will introduce our recent work on a conflict-driven scheduling framework to guarantee exactness while being scalable and versatile with a rich set of design constraints. I will also discuss our work on dynamically scheduled HLS to handle design behaviors not known at compile time. I conclude by outlining some of the new research directions being pursued by my group to further democratize heterogeneous accelerator-rich computing.
Bio
Zhiru Zhang is an assistant professor in the School of ECE at Cornell University and a member of the Computer Systems Laboratory. His current research focuses on high-level design automation for heterogeneous computing. His work has been recognized with a best paper award from TODAES (2012), the Ross Freeman award for technical innovation from Xilinx (2012), an NSF CAREER award (2015), a DARPA Young Faculty Award (2015), the IEEE CEDA Ernest S. Kuh Early Career Award (2015). He co-founded AutoESL Design Technologies, Inc. to commercialize his PhD research on high-level synthesis. AutoESL was acquired by Xilinx in 2011 and the AutoESL tool was rebranded as Vivado HLS after the acquisition.
Systems across the computing spectrum, from edge devices to cloud datacenters, are increasingly turning to specialized hardware accelerators for improved performance and energy efficiency. However, greater performance-per-watt comes at the cost of a much higher development effort. As the practice of traditional register-transfer-level (RTL) design has become unequivocally difficult, if not already unsustainable, high-level synthesis (HLS) has emerged as a promising approach to productive hardware specialization by enabling automatic generation of cycle-accurate RTL from untimed functional descriptions.
In this talk, I will review the progress we have made on HLS in recent years. Several real-life applications will be used as case studies to show the benefits and limitations of the current HLS tools. I will then mainly focus my talk on scheduling, which forms the algorithmic core to HLS. I will first describe a family of novel static scheduling algorithms devised for achieving high quality-of-results out-of-the-box. In particular, I will introduce our recent work on a conflict-driven scheduling framework to guarantee exactness while being scalable and versatile with a rich set of design constraints. I will also discuss our work on dynamically scheduled HLS to handle design behaviors not known at compile time. I conclude by outlining some of the new research directions being pursued by my group to further democratize heterogeneous accelerator-rich computing.
Bio
Zhiru Zhang is an assistant professor in the School of ECE at Cornell University and a member of the Computer Systems Laboratory. His current research focuses on high-level design automation for heterogeneous computing. His work has been recognized with a best paper award from TODAES (2012), the Ross Freeman award for technical innovation from Xilinx (2012), an NSF CAREER award (2015), a DARPA Young Faculty Award (2015), the IEEE CEDA Ernest S. Kuh Early Career Award (2015). He co-founded AutoESL Design Technologies, Inc. to commercialize his PhD research on high-level synthesis. AutoESL was acquired by Xilinx in 2011 and the AutoESL tool was rebranded as Vivado HLS after the acquisition.