CUDA Toolkit 13.4 Brings Windows on Arm Support, Rubin Preview, MPS V3

NVIDIA has released CUDA Toolkit 13.4, extending official support to Windows on Arm and introducing early developer access to the upcoming Rubin GPU architecture as a compute capability 107 preview. The release also ships Multi-Process Service (MPS) V3, a modernized control layer for shared GPU resources featuring a scriptable CLI, TOML configuration, SM partitioning controls, and cgroup-integrated memory limits for containerized workloads.

Beyond platform support, 13.4 adds CUDA Compute Fabric Transport for low-level NVLink data movement, programmatic access to locality domains for co-locating compute and memory, and new APIs for querying unified memory residency. NVIDIA has also decoupled the CUDA driver from the toolkit installer and switched coherent platforms like Grace Hopper and Grace Blackwell to Coherent Driver-based Memory Management by default.

For Python developers, cuda.core 1.1.0 introduces texture and surface programming APIs, NUMA-aware managed memory controls, and improved CUDA graph integration, while NVCC adds support for GCC 16, Clang 22, and the new SM_107 target for Rubin GPUs. Together these changes give engineers finer-grained control over multi-tenant GPU environments and prepare toolchains for next-generation hardware.