CUDA Toolkit 13.4 Adds Windows on Arm Support and Greater Control over Shared GPUs

The CUDA toolkit is still the choke point between new NVIDIA silicon and everything developers actually ship. Version 13.4 extends the platform map to Windows on Arm, previews the Rubin architecture, and hardens how teams carve up busy GPUs in containers.

For anyone building agentic systems, training stacks, or multi-tenant inference, those are not changelog trivia. They decide which laptops can develop, which racks can isolate noisy neighbors, and how early software can meet the next GPU generation.

We’re constructive on this release because it reads like platform work, not vapor. Windows on Arm support finally mirrors a capability Linux-on-Arm developers have had for a long time, which matters as Arm laptops and edge boxes show up in real AI workflows. Preview functional support for Rubin at compute capability 107 lets porting start before general availability—useful if teams treat preview seriously and do not confuse it with production readiness.

MPS V3 is the sleeper for shared clusters. A scriptable CLI, named server instances, TOML config, SM partition controls, and cgroup-tied memory limits are how you stop multi-process GPU sharing from being folklore. In containerized shops, that is the difference between “utilization” and “isolation.” CUDA Compute Fabric Transport aims at communication-library authors who need NVLink fabric moves without mapping every remote allocation into a process VA space—niche, powerful, easy to misuse if app developers skip NCCL or NVSHMEM.

Python and CCCL updates look like the daily-driver improvements. cuda.core gains texture/surface APIs, NUMA-aware managed memory knobs, and type stubs agents can read; cuda.compute adds ahead-of-time multi-arch algorithm artifacts. CCCL’s faster Blackwell DeviceScan and single-call CUB overloads cut boilerplate for people who live in device-wide primitives. Nsight Python 1.0 trying to make architectural profiling a decorator instead of a scavenger hunt is overdue.

The trajectory if this sticks: Windows-on-Arm AI development stops being a second-class path, Rubin software readiness starts earlier, and shared-GPU ops get closer to first-class policy. The downside is complexity—more transport APIs, memory modes, and preview targets for teams to misconfigure. Installers no longer bundling the driver also means upgrade discipline gets stricter.

CUDA 13.4 also exposes locality domains, unified-memory residency queries, GCC 16/Clang 22 host support, coherent-platform CDMM defaults, and a wide Nsight Systems/Compute pass including Rubin and Windows on Arm coverage. Core math libraries gain Rubin functional support and Windows-on-Arm paths for the N1X laptop ecosystem.