AI at Exascale
RAISE works on the convergence of traditional HPC and innovative AI techniques, thus leading us into an exciting time for accelerating scientific discovery and advancing engineering powered by an unprecedented Hardware Infrastructure (1) in the Exascale era. Based on that strong foundation, a seamlessly usable and versatile Software Infrastructure (2) is critical for accelerating convergence through new AI toolsets that are ready to scale for enormous quantities of datasets. RAISE considers AI requirements of Computing-driven Use Cases (3) using numerical methods based on known physical laws on the one hand and addressing AI requirements of Data-driven Use Cases (4) with large datasets of measurement devices on the other hand. 'AI at Exascale' in RAISE means to develop Unique AI Framework (5) methodologies co-designed by the above use cases but is usable by a wide variety of scientific and engineering applications in the Exascale era.
It addresses the needs of a broad range of HPC workloads intertwined with complementary AI training and inference using advanced AI algorithms such as Gated Recurrent Units, Residual Networks, Deep Belief Networks, and Stacked Auto-Encoders, to list just a few. Using data augmentation techniques and pre-trained neural networks through transfer learning, the RAISE AI framework further enables new approaches for scientific and engineering applications while saving valuable computational processing time in the Exascale era. Based on open source community toolsets and HPC best practices for AI, the framework will provide methodologies to scale-up AI on a large number of GPUs to keep pace with the ever-increasing number of applications that require Exascale computing. To tackle the fundamental problem of using AI in practical applications, w.r.t. hyper-parameter tuning requirement, the RAISE framework will leverage cutting-edge Neural Architecture Search techniques driven by reinforcement learning and evolutionary computation methods.
​
Offering this unique AI framework on emerging EuroHPC Pre-Exascale and Exascale HPC machines and providing support for its uptake by end-user communities of the Partnership for Advanced Computing in Europe (PRACE) represents an amplification factor for its use in scientific and engineering domains. Several National Competence Centers (NCCs) of the EuroCC project involved in RAISE will further raise awareness of the framework's unique AI at Exascale capabilities and foster its long-term sustainability.