NVIDIA has extended the reach of its CUDA Toolkit to Windows on Arm systems, bringing GPU-accelerated computing to a new environment previously limited to Linux. CUDA Toolkit 13.4 adds functional support for the NVIDIA Rubin architecture (compute capability 107) as a preview, enabling developers to begin porting applications before CUDA support for Rubin reaches general availability in a future release of the CUDA Toolkit, the company says.
The update introduces Multi-Process Service V3, providing developers with a scriptable command line interface and namespaces for “precise GPU partitioning, where compute performance, memory boundaries, and execution priority are defined programmatically,” according to NVIDIA documentation.
CUDA 13.4 Support for Windows on Arm Systems
Expanding the CUDA ecosystem, NVIDIA now supports application development on Windows on Arm platforms, a capability previously limited to Linux environments. This addition broadens the reach of GPU-accelerated computing to a different Windows-based environment, allowing developers to use NVIDIA’s tools across a wider range of devices and operating systems. The latest toolkit introduces Multi-Process Service V3, a modernized control layer for CUDA MPS, simplifying the automation and management of shared GPU resources. This new iteration moves beyond basic GPU sharing, offering developers a scriptable command line interface and named server instances.
API support for querying residency information for unified memory gives performance-sensitive libraries and runtimes a direct way to understand where managed or system-allocated data currently resides. With residency queries, CUDA applications and libraries make smarter decisions about where and when to schedule computations and data movement. Updates to core math libraries, including cuBLAS, further enhance performance. Functional support for the Rubin GPU architecture and Windows on Arm is also available for the N1X Laptop ecosystem.
These library updates, combined with the new MPS features and early Rubin access, position CUDA Toolkit 13.4 as a comprehensive upgrade for developers seeking to maximize the potential of NVIDIA GPUs. NVIDIA Nsight Systems 2026.5.1 expands platform coverage and workload visibility, now supporting CUDA 13.4, Rubin GPUs, and Windows on Arm.
The web release projects the NVTX range into an “All Streams” hierarchy, demangles cuTile names, and displays CUDA workloads submitted through CiG streams on the timeline. These enhancements provide developers with deeper insights into application performance, enabling more effective debugging and optimization. The inclusion of support for the latest NVIDIA hardware and software features ensures that Nsight Systems remains a valuable tool for developers working on advanced projects.
NVIDIA Rubin Architecture Preview via CUDA Toolkit 13.4
The release also introduces enhanced GPU management capabilities, expanded CUDA Python and CCCL functionality, and updates across NVIDIA Nsight developer tools and core math libraries. A key enhancement within CUDA Toolkit 13.4 is the introduction of residency queries, allowing CUDA applications and libraries to make informed decisions regarding computation scheduling and data movement.
The decoupling of the NVIDIA driver from the CUDA Toolkit installer represents a shift in distribution strategy, according to the company. Previously bundled together, the driver and toolkit are now separate installations, allowing users to select and manage drivers independently using their preferred package manager. On NVIDIA’s coherent platforms, including Grace Hopper, Grace Blackwell, and Vera Rubin, the driver now defaults to Coherent Driver-based Memory Management (CDMM) instead of NUMA.
NUMA mode remains fully supported and can be selected with a kernel module parameter. Updates extend beyond core functionality to include CUDA Python, which expands Pythonic access to CUDA APIs and algorithms. The latest version, cuda.core 1.1.0, introduces stable Pythonic APIs for texture and surface programming, enhanced managed-memory control, and improved CUDA graph integration. These additions streamline development workflows for Python-based CUDA applications.
Multi-Process Service V3 Simplifies GPU Resource Management
The update also incorporates TOML configuration support, streaming multiprocessor partition controls, and integration with cgroup-integrated GPU memory limits, allowing for precise GPU partitioning, where compute performance, memory boundaries, and execution priority are defined programmatically. This release ensures that MPS integrates into containerized environments, maximizing hardware utilization while maintaining strict resource isolation for every process. To get started with MPS V3, see the quickstart and full documentation.
CUDA Compute Fabric Transport for Scalable GPU Communication
CUDA Compute Fabric Transport (CFT) introduces a new approach to data movement across NVIDIA NVLink fabric, designed for applications demanding extreme scalability in multi-GPU systems. Rather than allocating virtual address space for every remote GPU, CFT allows software to target logical endpoints with an endpoint ID and offset, initiating asynchronous put, get, and reduction operations directly from the GPU. This method significantly reduces virtual address pressure, a common bottleneck in large-scale GPU deployments, and supports both unicast and multicast communication patterns.
The architecture reports completion and error status, enabling applications to detect and recover from failed fabric transfers, which is important for maintaining reliability in complex computations. CFT is implemented through the CUDA Driver API and is specifically intended for developers of communication libraries requiring highly customized functionality beyond what is offered by higher-level tools like NVIDIA NCCL or NVSHMEM.
According to NVIDIA documentation, most application developers will continue to benefit from utilizing these established libraries. The design prioritizes flexibility for those building the foundational communication layers, allowing for optimization tailored to specific workloads. This transport-centric approach contrasts with traditional methods of GPU communication, which often rely on extensive memory mapping and synchronization. CFT’s asynchronous nature allows for overlapping computation and communication, potentially leading to substantial performance gains.
The ability to specify offsets within logical endpoints also enables fine-grained data access, reducing unnecessary data transfers and improving overall efficiency. CUDA 13.4 enhancements include CFT, along with enhancements to the CUB library, including a faster cub::DeviceScan on NVIDIA Blackwell GPUs, single-call APIs for device-wide algorithms, and batched warp reductions. These improvements streamline the development process and simplify the implementation of complex algorithms.
A new kernel benchmarking decorator and context manager also provide a streamlined method for architectural metric collection, GPU-throttling prevention, and performance visualization, eliminating the need for manual report parsing. The toolkit’s advancements extend to debugging and profiling as well. Nsight Compute 2026.3 adds Tile IR support for CUDA Tile workloads, enabling developers to inspect Tile IR in the source page and correlate it with CUDA Tile source and generated code.
Compute Sanitizer now features improved shared memory out-of-bounds detection with compile-time patching on Hopper and newer architectures, enhancing code reliability. These updates, combined with the introduction of CFT, demonstrate NVIDIA’s ongoing efforts to provide developers with a comprehensive and powerful toolkit for building high-performance, scalable applications, the firm reports.
The company’s commitment to innovation is further underscored by its partnerships with organizations like QuEra Computing and Classiq, integrating their technologies with the CUDA platform to accelerate quantum-classical workflows. The launch of NVQLink in January 2026, an open architecture for low-latency integration of quantum processors with GPU supercomputers, exemplifies this collaborative approach.
Locality Domains and Unified Memory Residency Queries Enhance Performance
The latest CUDA Toolkit release introduces programmatic access to locality domains, discrete portions of a GPU containing streaming multiprocessors and device memory, allowing applications to allocate memory and resources within a specific area. This granular control enables co-location of computation with the memory it accesses, a strategy NVIDIA asserts can improve performance on devices featuring multiple locality domains. Developers can now define where data resides and processing occurs, potentially reducing latency and maximizing throughput by minimizing data transfer distances within the GPU.
For detailed implementation guidance, NVIDIA directs users to the CUDA Programming Guide. Beyond spatial optimization, CUDA Toolkit 13. NVIDIA highlights that access to this information is detailed in the API reference for cudaMemGetLocationInfo.
See today’s quantum computing news on Quantum Zeitgeist for the latest breakthroughs in qubits, hardware, algorithms, and industry deals.
