CUDA Toolkit 13.4 adds Windows on Arm support and better shared GPU management

NVIDIA has announced the latest version of the CUDA Toolkit 13.4, which not only expands platform support, but also gives developers comprehensive tools to manage shared GPU resources. The main highlight of the new features is official support for Windows on Arm (WoA), which until now was only available through the Linux environment. This capability allows developers to start porting CUDA applications to Windows Arm64 devices, check dependencies, and test the performance of CUDA applications on existing WoA systems. This is an important step towards the upcoming NVIDIA RTX Spark platform, which is planned for release in the near future.

In addition to platform enhancements, CUDA Toolkit 13.4 introduces early support for the Vera Rubin GPU architecture (compute capability 107). This architecture, described by NVIDIA as the next generation of agent-based AI, is now available to developers as a preview feature. This allows developers to begin optimizing and porting their applications to the Rubin architecture before it is officially supported in future CUDA releases.

Windows on Arm support – what does it mean for developers?

NVIDIA says that CUDA applications have long been able to run on Arm platforms via Linux, but version 13.4 extends this capability to Windows on Arm. This is important not only because of the growing market share of Windows Arm devices, but also because of the upcoming RTX Spark PCs that will be based on this architecture. Developers can now use NVIDIA tools to verify that their algorithms are working effectively and make the necessary performance improvements before the final product release.

MPS V3 - Modernized Shared GPU Management

CUDA Toolkit 13.4 introduces the third version of the Multi-Process Server (MPS) – MPS V3. This new control layer simplifies automated management of shared GPU resources. Key features include:

Scriptable CLI – a command line tool that allows you to automate MPS configuration.
Server instances and namespaces – allows you to organize parallel work and separate it logically.
TOML configuration support – simplifies the creation and management of configuration files.
SM (Streaming Multiprocessor) partition control – allows you to programmatically determine how many SM units are allocated to a specific process.
cgroup-integrated GPU memory limits – provides the ability to limit GPU memory usage in containers.

These capabilities allow for precise GPU partitioning policies: compute performance, memory limits, and execution priority can be set programmatically. In addition, MPS V3 integrates into containerized environments, ensuring maximum hardware utilization and strict process isolation.

CUDA Compute Fabric Transport (CFT) – a new data transfer model

A new feature, CUDA Compute Fabric Transport (CFT), provides a transport-centric way to move data across the NVIDIA NVLink fabric. Instead of the traditional mapping of remote GPU memory to the process„ virtual address space, CFT allows applications to use logical endpoints (endpoint IDs) and perform asynchronous “puts„, “gets”, and reduce operations directly from the GPU. This reduces virtual address pressure in large multi-GPU systems, supports unicast and multicast communication models, and provides error and completion status reporting that allows applications to detect, retry, or roll back failed transfers.

CFT is only available through the CUDA Driver API and is intended for developers of communication libraries who need specific features not available in higher-level libraries such as NCCL or NVSHMEM.

Locality Domain and Unified Memory Improvements

CUDA Toolkit 13.4 also opens up programmatic access to „locality domains,“ the parts of the GPU that contain the SM and device memory. Applications can allocate memory in a specific domain and create a „green context“ with SM resources in the same domain, reducing data transfer times and improving performance. In addition, an API has been introduced to query unified memory residency information, enabling libraries to make informed decisions about data location and movement.

Other improvements and implementation changes

In the new version, NVIDIA no longer includes the NVIDIA driver in the CUDA SDK installation package - users must install the "nvidia-open" driver separately or use their preferred package manager. Additionally, on coherent platforms such as NVIDIA Grace Hopper, Grace Blackwell, and Vera Rubin, the default driver now works with Coherent Driver-based Memory Management (CDMM) instead of the traditional NUMA mode, although NUMA is still supported and can be enabled via a kernel module parameter.

All of these innovations show that NVIDIA is working to not only expand the reach of the CUDA ecosystem, but also to give developers deeper control and flexibility to manage modern GPU architectures and resources.

Conclusions

CUDA Toolkit 13.4 brings significant improvements: Windows on Arm support, early access to the Rubin GPU architecture, improved MPS V3 management, and a new CFT data transport. These features allow developers to more efficiently build, test, and optimize applications for future platforms, including RTX Spark and agent-based AI. NVIDIA also emphasizes an open and modular implementation strategy that allows for the use of separate drivers and a choice of memory management models based on needs.

For developers already working with CUDA, these innovations will provide new opportunities to take advantage of modern GPU architectures and better manage shared resources in both physical and containerized environments.

Sources

IT SERVICES

Let's transform technology real results for your business.

We help companies apply artificial intelligence, automation, internet systems, and other digital solutions to real business processes.