Abstract: |
|
Three confluent factors have ignited the current AI revolution: availability of massive data, effective “deep learning” algorithms and models, and the wide availability of unprecedentedly-powerful and efficient processors in the form of CUDA-capable GPU accelerators. NVIDIA HW, SW, and accelerated library advances have increased neural network training performance by more than 65x in 3 years, supporting order-of-magnitude-per-year growth rates in deep neural network model size and complexity as well as similar increases in the size of data sets employed in training. Clearly HPC is shaping AI. How is AI shaping HPC? Beyond increased optimization for reduced precision computation, how is the need to continue to increase training performance influencing the architecture and implementation of GPUs, the most efficient programmable computational accelerator known today? Beyond the processor, how will AI-driven optimization alter traditional HPC node and system architectures, storage and networking, programming models and workflows, and the SW tools and system environments in everyday use at supercomputing sites around the world today? From the perspective of a computational scientist today — perhaps already reeling from the need to “modernize” their codes to maximize parallelism in the shadow of Moore’s Law’s slow demise — is this going to be a good thing or a bad thing? Using the ways AI is likely to be (or is already being) employed to augment or disrupt classical high-performance computing as context, this talk will consider these questions and propose plausible (or, at least, possible) projections for a future where HPC and AI synergistically reshape computational science, industry, economics, and — ultimately — society. |
|