Explore the world of general-purpose, task-based runtime systems in this 38-minute conference talk from NVIDIA GTC 2025. Learn how to write multi-GPU libraries that scale transparently to any number of GPUs without requiring complex explicit communication and coordination code. Discover techniques for making your code interoperate seamlessly with existing domain libraries within the same program. Presented by Manolis Papadakis, Software Engineering Manager of the Legate Framework at NVIDIA, this session (S72579) provides valuable insights for developers looking to leverage accelerated computing libraries and CUDA technology for multi-GPU scaling. Part of the "Models / Libraries / Frameworks - Accelerated Computing Libraries" topic track at GTC 2025.