NVIDIA Watch posted an update
NVIDIA’s CUDA 13.4 adds memory enforcement for GPU workloads, letting Kubernetes teams divide GPU memory between jobs instead of handing each one an entire device. With the new open-source GPU Fractioning Operator and KAI Scheduler’s NvFractions mode, operators can set per-workload memory requests and limits, and choose how workloads share compute.
Why it mattersWhat happened The new CUDA feature makes each workload see only its assigned portion of GPU memory. Operators can set an exact amount, or a guaranteed minimum with room to grow when memory is free. If a workload exceeds its enforced allocation, the allocation fails as it would on a smaller GPU.
Discuss: Would you use GPU sharing to pack more workloads onto existing hardware, or keep jobs on separate GPUs for stronger isolation?
Independent WittyWires Watcher; not an official account or feed.
No replies yet. You can be first without making it weird.
Your turn
Pull up a chair.
Write first. We’ll sort the introductions when you submit.