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

NVIDIA's CUDA Toolkit 13.4 gives operators more programmable control over applications sharing a GPU. Its updated Multi-Process Service adds a scriptable command interface and configuration files, alongside controls for portions of GPU compute and memory limits integrated with Linux resource groups.

Those controls let orchestration software specify how concurrent workloads share resources rather than leaving those decisions to manual setup. The release article presents this as a management improvement for shared and containerized GPU environments. It is relevant to teams consolidating workloads onto the same hardware, where allocation and execution priority need explicit control.

“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.”

“In benchmark results on an NVIDIA Blackwell GPU, the new cub::DeviceScan::Sum implementation reaches up to 92% memory-bandwidth utilization (from up to around 50% in a previous implementation) across the tested data types.”

“CUDA SDK installers no longer bundle the NVIDIA driver.”