Accelerating Deep Learning with Analog In-Memory Compute
Keynote 1 : Wednesday, June 1, 9:00 AM - 10:00 AM
Abstract: Artificial Intelligence (AI) is all around us today - augmenting our capabilities and enriching our experiences - but it was less than a decade ago that the first key breakthroughs in deep learning were made. Tremendous progress has been made since in expanding AI applications as well as the accuracy of AI models, often by developing larger models that are trained on larger datasets. However, this explosive growth in model size and the concomitant increase in required compute is unsustainable without significant innovations across the hardware stack. We will present a novel non-von Neumann computational approach that envisions artificial neural networks mapped to arrays of non-volatile memory (NVM) elements. These NVM elements act as artificial synapses and encode the weights of a neural network that execute compute operations in-memory and constant time, thereby enabling significant power performance benefits. Analog NVM synapses are attractive given their non-volatility, analog tunability and maturity of the technology. This talk details innovations across materials, algorithms, and architecture to enable acceleration of AI inference and training jobs. Indeed, novel compute paradigms combined with heterogeneous integration technologies that address key bandwidth & connectivity challenges will be needed to enable energy efficiency improvements by orders of magnitude to power the AI of tomorrow.
Biography: Dr. Vijay Narayanan received his B.Tech. in Metallurgical Engineering from the Indian Institute of Technology, Madras (1995), and his M.S. (1996) and Ph.D. (1999) in Materials Science and Engineering from Carnegie Mellon University. After completing post-doctoral research at Arizona State University, Dr. Narayanan joined the IBM T. J. Watson Research Center in 2001 where he pioneered High-k /Metal Gate Research and Development from the early stages of materials discovery to development and implementation in manufacturing. These High-k/Metal Gate materials form the basis of all recent IBM systems processors and of most low-power chips for mobile devices. Currently, Dr. Narayanan is an IBM Fellow and Senior Manager at IBM Research where he is the strategist for Physics of AI and leads a worldwide IBM team developing Analog Accelerators for AI applications within the IBM Research AI Hardware Center. Dr. Narayanan is an IEEE Senior Member and was elected a Fellow of the American Physical Society in 2011. He is an author of over 100 journal and conference papers, holds more than 230 US patents, and has edited one book: "Thin Films On Silicon: Electronic And Photonic Applications".
Thermal Management Challenges for Battery Electric Vehicles
Keynote 2 : Thursday, June 2, 9:00 AM - 10:00 AM
Abstract: In recent years, several key external drivers and technology trends have greatly accelerated the move towards electrification for on road vehicle propulsion. These drivers include societal concerns for climate change, the ongoing push for continuous improvement of air quality, rapid expansion of renewable energy sources and the emergence of vehicle connectivity. In response, Ford has greatly increased its investments in vehicle electrification and is moving aggressively into the development and deployment of battery electric vehicles across its product portfolio. Battery electric propulsion systems feature highly efficient components that present unique thermal management challenges compared with established internal combustion engine propulsion systems. This talk will discuss several of these critical challenges particularly with respect to battery thermal management and related interactions with other subsystems (e.g. cabin climate control) during cold and hot ambient vehicle operation. The associated system design trade-offs and impacts to key vehicle level attributes will be highlighted including the importance to overall vehicle energy management, range and performance. Key differences across the spectrum of vehicle applications and the significant emerging role of vehicle connectivity will also be discussed.
Biography: Mark Jennings is Senior Technical Leader for Vehicle Energy Management & Propulsion Systems Analysis for Ford Motor Company working in Ford's Research & Advanced Engineering organization on electrified powertrain systems. Mark's work at Ford on electrified powertrain systems has covered a range of system technologies encompassing mild/medium hybrid electric vehicle (HEV) systems, full HEV systems, plug-in HEV systems, battery electric vehicles and fuel cell electric vehicles. Throughout his years at Ford, he has led numerous efforts to define and assess new electrified powertrain system concepts. Through this work he has played a significant role in defining Ford's powertrain electrification strategy. He has also been a leader in establishing and applying model-based development and optimization methodologies towards the advancement of electrification technologies for vehicle propulsion. Mark has BS, MS and PhD degrees in Mechanical Engineering.
Hyperdimensional Computing System Design & Thermal Management
Keynote 3 : Friday, June 3, 9:00 AM - 10:00 AM
Abstract: In today's world technological advances are continually creating more data than what we can cope with. Much of data processing will need to run at least partly on devices at the edge of the internet, but training state of the art neural networks at the edge is too costly. Hyperdimensional (HD) computing is a class of light-weight learning algorithms that is motivated by the observation that the human brain operates on a lot of data in parallel. HD computing uses high dimensional random vectors (e.g. ~10,000 bits) to represent data, making the model robust to noise and HW faults. It uses search, along with three base operations: permutation, addition (or bundling/consensus sum) and multiplication (circular convolution / XOR). Addition allows us to represent sets, multiplication expresses conjunctive variable binding, and permutation enables encoding of causation and time series. Hypervectors are compositional - they enable computation in superposition, unlike standard neural representations. Systems that use HD computing to learn can be accelerated directly in memory and storage and have been shown to be accurate, fast and very energy efficient. Most importantly, such systems can explain how they made decisions, resulting in devices that can learn directly from the data they obtain without the need for the cloud. In this talk I will present some of my teamâ€™s recent work on hyperdimensional computing theory, software and hardware infrastructure, including: i) novel algorithms supporting key cognitive computations in high-dimensional space such as classification, clustering, regression and others, ii) hardware acceleration of HD computing on GPUs, FPGAs, in memory and storage, along with software infrastructure to support it, iii) thermal management strategies to address thermal issues that arise when learning algorithms, such as HD computing, are accelerated in memory and storage.
Biography: Tajana Simunic Rosing is a Full Professor, a holder of the Fratamico Endowed Chair, ACM & IEEE Fellow, and a director of System Energy Efficiency Lab at UCSD. Her research interests are in energy efficient computing, computer architecture, distributed and embedded systems. She is leading a number of projects, including efforts funded by DARPA/SRC JUMP CRISP program, with focus on design of accelerators for analysis of big data including machine learning, image/video processing and bioinformatics; DARPA, NSF and SRC funded projects on Hyperdimensional Computing, SRC funded project acceleration of 3rd generation Fully Homomorphic Encryption, and NSF AI TILOS center projects on federated learning and AI-based chip design. She recently headed the effort on SmartCities that was a part of DARPA and industry funded TerraSwarm center. Tajana led the energy efficient datacenters theme in MuSyC center, and a number of large projects funded by both industry and government focused on power and thermal management. From 1998 until 2005 she was a full time research scientist at HP Labs while also leading research efforts at Stanford University. She finished her PhD in EE in 2001 at Stanford, concurrently with finishing her Masters in Engineering Management. Her PhD topic was dynamic management of power consumption. Prior to pursuing the PhD, she worked as a senior design engineer at Altera Corporation.