NVIDIA CUDA 13.5 with Vera Rubin support and more

General Availability for the Vera-Rubin Architecture

The transition to general availability means that the Vera-Rubin hardware is now fully supported for production environments. This release builds on the early support that was already included in previous CUDA 13 versions. In addition to official support status, CUDA 13.5 introduces specific performance optimizations and architectural enhancements designed to fully leverage the capabilities of the Vera Rubin hardware.

Multi-GPU Communication and Memory Optimization

CUDA 13.5 introduces several technical enhancements aimed at optimizing interaction and data exchange between multiple GPUs. The update improves multi-GPU communication and ensures more comprehensive interoperability with existing CUDA libraries through the integration of CUDA Tile.

For developers using CUDA Python, this version includes the implementation of bundled memory copies. This feature is designed to optimize the efficiency of data transfers between the host and the device, thereby reducing overhead in Python-based AI and data science workflows.

Resource Management and AI Profiling

The update expands the toolkit for managing and analyzing GPU workloads. New GPU resource management features have been added to enable finer control over hardware allocation. NVIDIA has also updated Nsight with new profiling capabilities. These tools were specifically designed to monitor and optimize AI training and inference processes, enabling developers to identify bottlenecks and improve the efficiency of large-scale model deployments.

Detailed technical specifications and complete developer notes for the stable CUDA 13.5 update can be found on the NVIDIA Developer Blog.