NVIDIA has released CUDA Toolkit 13.4, bringing CUDA support to Windows on Arm and giving developers preview functional support for its upcoming Rubin GPU architecture. The update also overhauls shared-GPU management and expands NVIDIA’s programming tools, libraries, and runtime APIs.
CUDA applications were already supported on Arm through Linux, but version 13.4 extends that capability to existing Windows-on-Arm systems. NVIDIA says developers can now port Windows Arm64 applications, validate dependencies, and test CUDA software paths before RTX Spark laptops launch. Wccftech reports that those laptops are expected to arrive in October.
The practical benefit is preparation time: developers do not need to wait for RTX Spark hardware before beginning compatibility work. NVIDIA has not specified which Windows versions or Windows-on-Arm devices are supported, so developers will need to consult the toolkit documentation when choosing test systems.
CUDA 13.4 also provides preview functional support for Rubin, listed as compute capability 107. This early implementation is designed for porting, developing, and validating applications for Vera Rubin CPUs and GPUs. Full CUDA support for Rubin will reach general availability in a future toolkit release, although NVIDIA has not announced when that will happen.
Shared GPU environments receive a substantial operational update through Multi-Process Service V3. Its modernized control layer includes a scriptable command-line interface, named server instances, namespaces, and TOML configuration. Streaming-multiprocessor partition controls and cgroup-integrated GPU-memory limits enable programmatic partitioning, including within containerized environments.
Elsewhere, CUDA Compute Fabric Transport adds a transport-focused API for moving data across NVIDIA NVLink fabric. Advanced applications and communications libraries can use named logical endpoints alongside asynchronous put, get, and reduction operations. CUDA 13.4 also updates the compiler, PTX ISA, CUDA C++, CUDA Tile programming, runtime APIs, core math libraries, and Nsight tools, while expanding CUDA Python and CCCL functionality.
The headline additions point in two directions: getting Windows-on-Arm developers working before RTX Spark arrives, and letting teams begin Rubin porting before that architecture receives full CUDA support. CUDA 13.4 puts both preparation tracks into developers’ hands now.
