Nvidia has described integrating CUDA 13.4 memory limits with a GPU Fractioning Operator and KAI Scheduler on Kubernetes. Workloads can receive guaranteed memory and a separate maximum, borrow unused capacity and fail allocations that exceed the limit. Time-slicing provides more favorable fault separation, while SM sharing uses a shared fault domain and is suited to mutually trusted workloads. No overall throughput or cost-saving percentage was supplied, so memory control should not be mistaken for a universal performance or security-isolation guarantee.