Title: Designing High Performance Software by Harnessing the Power of Hardware.

Abstract: The core computing kernels of important real-world applications are often computationally intensive. Well-design software that implements such a kernel requires a thorough understanding of issues transcending hardware and software boundaries to push the system to deliver the maximum performance for heightened end-user experience. This talk will focus on the challenges involved in designing two such real-world applications, highlighting the issues faced in their design and outlining solutions. The first application is a video encoding software for the next generation compression standard. The 5X increase in compute density of the algorithm over the previous-generation standard presents a significant challenge in implementing a competent encoder that may displace existing encoders without a significant increase in hardware cost. The second application is an automated audio-video sync detection engine that uses neural networks to mark out sections of the video that may be out-of-sync. Owing to the high demands on computation and memory placed by CNNs today, designing such an application to operate at real-time speeds on embedded systems presents challenges that warrant innovative solutions.