MONDAY October 14, 9:00am - 10:00am | Eisner Lubin Auditorium
EVENT TYPE: KEYNOTEHigh Performance Computing in a World of Embedded Intelligence
Steve Keckler - NVIDIA
The confluence of high-performance computing and massive training data sets has enabled machine learning algorithms to play a transformational role in many disciplines. While tremendous effort has been applied to both training and inference in datacenter settings, the world of embedded computing devices demands capabilities to perceive entities in a complex environment, understand their surroundings, and make intelligent decisions based on observations. Applications such as intelligent video analytics, autonomous vehicles, and robotics all have tremendous computational requirements, limited power budgets, and safety/privacy standards that must be met. This talk will describe requirements of these embedded intelligence workloads and opportunities for architecture and hardware/software co-design to provide the computational throughput within the requisite power envelopes. It will also discuss challenges for safety-critical systems and methods of understanding and mitigating such vulnerabilities. Finally, the talk will present an automated design methodology for customizing an architecture to the specific parameters of a machine learning application, enabling quick development of optimized ASIC designs.