CUDA Toolkit 13.4 Adds Windows on Arm Support and Rubin Preview

NVIDIA released CUDA Toolkit 13.4 on September 9, 2026, adding Windows on Arm support and a preview of the Rubin GPU architecture. The toolkit now lets developers build CUDA applications for Windows on Arm, extending the existing Linux on Arm support. The Rubin preview, with compute capability 107, allows early porting before general availability in a future release.

Shared GPU Control with MPS V3

Multi-Process Service V3 introduces a modernized control layer for managing shared GPUs. It provides a scriptable command-line interface, named server instances, and namespaces to organize concurrent workloads. TOML configuration files, streaming multiprocessor partition controls, and cgroup-integrated GPU memory limits enable programmatic definition of compute performance, memory boundaries, and execution priority. This allows precise GPU partitioning in containerized environments, maximizing hardware utilization while maintaining strict resource isolation. The release includes a quickstart guide and full documentation.

CUDA Compute Fabric Transport

CUDA Compute Fabric Transport (CFT) offers a transport-centric API for communication-library developers to move data across NVLink fabric. Instead of mapping every remote GPU allocation into a process's virtual address space, software can target named logical endpoints using an endpoint ID and offset, then issue asynchronous put, get, and reduction operations directly from the GPU. This reduces virtual-address pressure in large multi-GPU systems, supports unicast and multicast patterns, and reports completion and error status for retry or rerouting. CFT is available only through the CUDA Driver API and is intended for developers needing specific functionality not available in higher-level libraries like NCCL or NVSHMEM.

Locality Domains and Unified Memory Queries

CUDA 13.4 exposes programmatic access to locality domains, which are portions of a GPU containing streaming multiprocessors and device memory. Applications can allocate memory in a locality domain and create a green context with SM resources in the same domain, potentially improving performance by co-locating computation near its data. The release also adds API support for querying unified memory residency, giving performance-sensitive libraries a direct way to know where managed or system-allocated data resides. This helps avoid unnecessary page migrations and remote memory accesses.

Driver and Compiler Updates

The CUDA SDK installers no longer bundle the NVIDIA driver; developers must install the nvidia-open driver or cuda-toolkit packages separately. On coherent platforms like Grace Hopper, Grace Blackwell, and Vera Rubin, the driver now defaults to Coherent Driver-based Memory Management (CDMM) instead of NUMA, though NUMA remains supported via a kernel module parameter. Host compiler compatibility now includes GCC 16 and Clang 22, and the new SM_107 target enables compilation for Rubin GPUs.

CUDA Python Enhancements

cuda.core 1.1.0 adds texture and surface programming with first-class Python APIs. The new cuda.core.texture module includes OpaqueArray and MipmappedArray for hardware-laid-out GPU allocations, plus TextureObject and SurfaceObject for bindless, hardware-filtered kernel reads and typed kernel-side loads and stores. ManagedMemoryResource.allocate() now returns a ManagedBuffer with property-based interfaces for CUDA memory advice, including read-mostly data, preferred placement, and processor access. The new Host type complements Device for specifying memory locations, including NUMA nodes. The release also provides .pyi type stubs for every public API, improving IDE autocompletion and coding agent access. GraphBuilder.graph_definition exposes captured graphs as GraphDefinition, enabling combination of stream capture with explicit graph construction.

cuda.compute 1.1 enables ahead-of-time compilation of algorithm objects for multiple GPU architectures, including on build systems without a GPU. ProxyArray and ProxyValue describe argument types without allocating device memory, and serialize() creates artifacts for later deployment. On the target system, deserialize() restores the algorithm without recompiling and loads the build matching the current GPU architecture.

CCCL 3.4 delivers a faster warp-specialized cub::DeviceScan on Blackwell GPUs, reaching up to 92% memory-bandwidth utilization. It also adds single-call APIs across CUB device-wide algorithms, batched warp reductions, and C++ Standard Library parallel algorithms in cuda::std.