Keynote Speakers

Abstract
The increasing use of large-scale AI and deep learning models for various applications raises interest in efficient storage and process of model data.
Server systems for fast processing of vast amounts of model data usually use high-capacity DRAMs. However, high power consumption and cost often limit the scalability of system configuration. And, the overhead associated with data transmission between storage and DRAM is also another problem to avoid.

This presentation introduces computational storage which provides data processing capability, and explains the system configuration which effectively processes large-capacity model data stored inside the storage. These computational storage devices can be applied not only to large-scale AI applications but also to graph neural networks and recommended applications, which is expected to have advantages over existing AI systems using lagacy storage devices in terms of total cost of ownership. Samsung aims to discover the possibility of developing new applications through collaboration with various universities/companies using the new storage model.

Abstract
Modern-day AI has only been possible due to advances in semiconductor manufacturing leading to ever more powerful processors. However, solving today’s AI problems is a data center scale problem and powerful chips alone is not enough. The modern AI data center requires high performance compute, networking, and storage and even more importantly high performance software. NVIDIA’s AI software is used today by every semiconductor manufacturer. In his talk, Hamilton will address some of the most promising application of AI in semiconductor industry.

Abstract
The true potential of AI rests on super-human learning capacity, and on the ability to selectively draw on that learning. Both of these properties – scale and selectivity – challenge the design of AI computers and the tools used to program them. A rich pool of new ideas is emerging, driven by a new breed of computing company, according to Graphcore co-founder Simon Knowles. In his talk, Simon discusses the creation of the Intelligence Processing Unit (IPU) – a new type of processor, specifically designed for AI computation. He looks ahead, towards the development of AIs with super-human cognition, and explores the nature of computation systems needed to make powerful AI an economic everyday reality.

Abstract
Despite decades of progress in semiconductor scaling, computer architecture, and artificial intelligence, our computing technology today still lags biological brains in many respects. While deep artificial neural networks have provided breakthroughs in AI, these gains come with heavy compute and data demands relative to their biological counterparts. Neuromorphic computing aims to narrow this gap by drawing inspiration from the form and function of biological neural circuits. The past several years have seen significant progress in neuromorphic computing research, with chips like Intel’s Loihi demonstrating, for the first time, compelling quantitative gains over a range of workloads—from sensory perception to data efficient learning to combinatorial optimization. This talk surveys recent developments in this endeavor to re-think computing from transistors to software informed by biological principles. It previews a new class of chips that can autonomously process complex data streams, adapt, plan, behave, and learn in real time at extremely low power levels.