Stanford Seminar - New Golden Age for Computer Architecture - John Hennessy

EE380: Computer Systems Colloquium Seminar New Golden Age for Computer Architecture: Domain-Specific Hardware/Software Co-Design, Enhanced Security, Open Instruction Sets, and Agile Chip Development Speaker: John Hennessy, 2017 Turing Award Recipient / Chairman, Alphabet In the 1980s, Mead and Conway democratized chip design and high-level language programming surpassed assembly language programming, which made instruction set advances viable. Innovations like RISC, superscalar, multilevel caches, and speculation plus compiler advances (especially in register allocation) ushered in a Golden Age of computer architecture, when performance increased annually by 60%. In the later 1990s and 2000s, architectural innovation decreased, so performance came primarily from higher clock rates and larger caches. The ending of Dennard Scaling and Moore's Law also slowed this path; single core performance improved only 3% last year! In addition to poor performance gains of modern microprocessors, Spectre recently demonstrated timing attacks that leak information at high rates. We're on the cusp of another Golden Age that will significantly improve cost, performance, energy, and security. These architecture challenges are even harder given that we've lost the exponentially increasing resources provided by Dennard scaling and Moore's law. We've identified areas that are critical to this new age. Turing lecture presented at 2018 ACM/IEEE 45th Annual International Symposium on Computer Architecture (ISCA) and published in the Proceedings. About the Speaker: John L. Hennessy, Professor of Electrical Engineering and Computer Science, served as President of Stanford University from September 2000 until August 2016. In 2017, he initiated the Knight-Hennessy Scholars Program, the largest fully endowed graduate-level scholarship program in the world, and he currently serves as Director of the program. His honors include the 2012 Medal of Honor of the Institute of Electrical and Electronics Engineers and the ACM Turing Award (jointly with David Patterson). He is an elected member of the National Academy of Engineering, the National Academy of Science, the American Academy of Arts and Sciences, The Royal Academy of Engineering, and the American Philosophical Society. Hennessy earned his bachelor's degree in electrical engineering from Villanova University and his master's and doctoral degrees in computer science from the Stony Brook University. For more information about this seminar and its speaker, you can visit https://ee380.stanford.edu/Abstracts/181010.html Learn more: http://bit.ly/WinYX5 0:00 Introduction 0:28 Outline 1:27 IBM Compatibility Problem in Early 1960s By early 1960's, IBM had 4 incompatible lines of computers! 3:33 Microprogramming in IBM 360 Model 4:47 IC Technology, Microcode, and CISC 10:32 Microprocessor Evolution • Rapid progress in 1970s, fueled by advances in MOS technology, imitated minicomputers and mainframe ISAS Microprocessor Wers' compete by adding instructions (easy for microcode). justified given assembly language programming • Intel APX 432: Most ambitious 1970s micro, started in 1975 15:12 Analyzing Microcoded Machines 1980s 17:40 From CISC to RISC . Use RAM for instruction cache of user-visible instructions 19:24 Berkeley & Stanford RISC Chips 20:06 "Iron Law" of Processor Performance: How RISC can win 22:59 CISC vs. RISC Today 25:19 From RISC to Intel/HP Itanium, EPIC IA-64 26:49 VLIW Issues and an "EPIC Failure" 29:18 Fundamental Changes in Technology 31:00 End of Growth of Single Program Speed? 33:28 Moore's Law Slowdown in Intel Processors 33:57 Technology & Power: Dennard Scaling 35:02 Sorry State of Security 36:35 Example of Current State of the Art: x86 . 40+ years of interfaces leading to attack vectors · e.g., Intel Management Engine (ME) processor . Runs firmware management system more privileged than system SW 40:33 What Opportunities Left? 41:59 What's the opportunity? Matrix Multiply: relative speedup to a Python version (18 core Intel) 43:14 Domain Specific Architectures (DSAs) • Achieve higher efficiency by tailoring the architecture to characteristics of the domain • Not one application, but a domain of applications 44:11 Why DSAs Can Win (no magic) Tailor the Architecture to the Domain • More effective parallelism for a specific domain 46:08 Domain Specific Languages 47:14 Deep learning is causing a machine learning revolution 48:27 Tensor Processing Unit v1 48:48 TPU: High-level Chip Architecture 49:55 Perf/Watt TPU vs CPU & GPU 50:34 Concluding Remarks