CUDA 13.4 adds Windows on Arm GPU sharing

NVIDIA released CUDA Toolkit 13.4, adding Windows on Arm support, preview support for Rubin at compute capability 107, and MPS V3 controls for partitioning shared GPU compute and memory. NVIDIA’s release says the toolkit also ships CCCL 3.4 and raises Blackwell DeviceScan utilization to up to 92%, from around 50% in its earlier implementation—a vendor benchmark, not a workload guarantee. GPU teams should test the upgrade on isolation and memory-bound workloads, alongside the archive’s PAIR routing analysis before treating shared-GPU controls as a capacity multiplier.