NVIDIA Unveils CUDA Toolkit 13.0 Enhancements for Jetson Thor

Lawrence Jengar
Sep 02, 2025 17:30

NVIDIA announces CUDA Toolkit 13.0 for Jetson Thor, featuring a unified Arm ecosystem, enhanced virtual memory, and improved GPU sharing, streamlining development for edge computing.

NVIDIA is set to revolutionize the world of embedded and edge computing with the release of CUDA Toolkit 13.0 for its Jetson Thor System on Chip (SoC). This update, powered by the NVIDIA Blackwell GPU architecture, promises to enhance speed, efficiency, and versatility, according to NVIDIA.

Unified CUDA Toolkit for Arm Platforms

The most significant change in CUDA 13.0 is the unification of the CUDA toolkit for Arm platforms, which eliminates the need for separate toolkits for server-class and embedded systems. This streamlining allows developers to build applications once and deploy them across various platforms without code changes, significantly boosting productivity.

Jetson Thor will now support Unified Virtual Memory (UVM) with full coherence, which allows the device to access pageable host memory via the host’s page tables. This update aligns Jetson platforms with discrete GPU systems in terms of UVM functionality.

Enhanced GPU Sharing Features

CUDA 13.0 introduces several GPU sharing features to improve utilization and performance. The Multi-Process Service (MPS) feature enables multiple processes to share the GPU concurrently, enhancing throughput and scalability without requiring changes to application code. This is particularly beneficial for applications with small or bursty workloads.

Additionally, the new green contexts feature allows for deterministic GPU scheduling by pre-assigning resources, ensuring predictable execution for latency-sensitive workloads. This is crucial for applications like robotics, where real-time performance is essential.

Developer Tool Enhancements

Certain developer tools, including the nvidia-smi utility and the NVIDIA Management Library (NVML), are now supported on Jetson Thor. These tools provide better insight into GPU usage and control over resources, although some features such as clock and power queries are expected in future updates.

Future Developments

Looking forward, NVIDIA plans to introduce the Multi-Instance GPU (MIG) feature, which will allow the partitioning of large GPUs into smaller, isolated devices. This will enable workloads with mixed criticality to run in parallel, improving determinism and fault isolation.

Developers can explore these new features in the CUDA 13.0 toolkit available in the JetPack 7.0 release, and NVIDIA encourages engagement through their developer forums as they continue to enhance the capabilities of the Jetson platform.

Image source: Shutterstock