CUDA adds enforced memory limits for shared GPUs
CUDA 13.4 adds enforced memory boundaries for workloads sharing an NVIDIA GPU, with fixed caps or a reserved minimum and burst limit. A GPU Fractioning Operator applies these settings, while KAI Scheduler assigns capacity in Kubernetes. Compute-sharing mode matters: time-sliced workloads have separate error isolation, but workloads running concurrently through SM sharing share an error domain and should be trusted together.
NVIDIA Technical Blog · Oct 27, 2026
Why it matters
GPU operators can enforce memory allocation limits while explicitly choosing the failure boundary between neighboring workloads.
More context
The documented stack combines the NVIDIA GPU Operator, the new fractioning operator and KAI Scheduler 0.18.0 in NvFractions mode. Start with a small allocation test and verify over-limit failures and the chosen compute mode before expanding.
Keep in mind
These software boundaries should not be assumed equivalent to hardware-isolated MIG partitions. The article provides mechanisms, not a workload-independent utilization or speed guarantee.
Original read