Cloud TPU Pods - AI Supercomputing for Large Machine Learning Problems
Explore the technical details of Cloud TPU and Cloud TPU Pod in this 42-minute conference talk from Google I/O'19. Dive into the domain-specific architecture designed to accelerate TensorFlow training and prediction workloads, providing performance benefits for machine learning production use. Discover new TensorFlow features enabling large-scale model parallelism for deep learning training. Learn about GPU V2, object detection, Google Cloud Platform Notebook, interconnect technology, pricing considerations, and scalability. Gain insights into model parallel techniques, training transformer models, and optimizing learning rate schedules. Presented by Kaz Sato and Martin Gorner, this talk covers topics such as accuracy boosting, data power, complexities in AI supercomputing, and includes a demo showcasing the capabilities of Cloud TPU Pods for large machine learning problems.