What's New in CUDA 13.4

CUDA 13.4 brings changes across the toolkit: new CUDA Tile features, new PTX instructions for kernel and library authors, locality-aware APIs for multi-die GPUs, and better GPU sharing for containerized workloads. It also includes a developer preview for the Rubin architecture, so you can target sm_107 and start porting your code before launch. In this livestream, the CUDA team walks through the features most likely to change how you write, tune, and deploy GPU code, with code examples and time for your questions. What you'll learn: - How programmatic dependent launch (PDL) works, and how to use it from cuTile Python and CUDA Tile IR to overlap a kernel’s setup with the end of the kernel before it. - What the new PTX instructions do, including L1 prefetch, int4-to-int8 matrix loads, packed FP8 arithmetic, and lighter-weight proxy fences, and which GPUs they target. - How to start porting to Rubin with the sm_107 developer preview, and how new locality APIs let you keep compute and memory on the same part of a multi-die GPU. - How cgroup-based GPU memory limits and the updated Multi-Process Service (MPS) make it safer to share a GPU across containers, plus updates to cuBLAS, NVCC, and the developer tools. Join us live and bring your questions for the CUDA team.