NVIDIA ships CUDA Toolkit 13.4 with Windows-on-Arm support and shared-GPU controls

NVIDIA released CUDA Toolkit 13.4, whose two headline additions are Windows-on-Arm support and greater control over shared GPUs, on top of routine functionality and performance improvements. Windows-on-Arm support matters as Arm-based Windows devices proliferate: developers can now target NVIDIA GPUs from that platform, widening CUDA's reach beyond x86 Windows and Linux.

The shared-GPU controls address a practical multi-tenant problem — letting operators partition and govern GPU resources more finely, which is increasingly important as inference workloads and dev environments contend for scarce accelerator capacity. This is developer-platform plumbing, but it's strategically consistent with NVIDIA's week: the Hugging Face acquisition is about widening the developer 'front door,' and CUDA remains the moat that keeps those developers on NVIDIA silicon.

Separately, NVIDIA published guidance on encode-prefill-decode (EPD) disaggregation for serving multimodal models — an inference optimization that splits the vision-encoder stage from prefill — reflecting its continued push into serving efficiency for vision-language workloads. The competitive context is the multi-accelerator pledge NVIDIA made around Hugging Face: expanding CUDA to new platforms is exactly the kind of lock-in that makes open-hub neutrality promises worth scrutinizing. Watch uptake of Windows-on-Arm CUDA among developers and whether the shared-GPU controls ease the capacity crunch teams are feeling across the industry.