Presentation

DescriptionArtificial intelligence, and more specifically neural-network based machine learning, has evolved in leaps and bounds, and is expanding from its original constituency of vision, perception and natural language processing into many scientific fields. At the same time, computational complexity and data volumes of AI applications (in particular of training phases) increasingly outgrow the capabilities of single-node systems, and optimizing end-to-end performance becomes a key issue ("Performance for AI") .

At the same time, the proven capability of AI techniques to analyze high-dimensional data, recognize complex patterns and solving difficult optimization problems promises to substantially improve the capabilities of large scale, parallel performance analysis and of automatic code tuning. ML-based analysis functionality has been integrated into parallel performance tools, and the role of AI-directed decisions in compilation and auto-tuning systems are active research topics ("AI for performance").

This BoF brings together expert AI technology and application developers with HPC application and tool developers and end-users to investigate status quo and opportunities both in "Performance for AI" and "AI for performance", and discuss how to accelerate mutual progress.

BoF Organizer/Speakers

Hans-Christian Hoppe

Principal Engineer

Intel

Intel Datacenter Group