10x Your Neocloud's GPU Utilization with HAMi

DRA gives Kubernetes a way to request fractional GPUs, but it does not solve the utilization problem by itself - the device still needs virtualization and a scheduler that knows how to share it. Hardware-agnostic GPU virtualization provides that layer: one scheduling layer shares NVIDIA, AMD, Ascend, and domestic NPUs under a single pool - fractional GPUs, mixed vendors, whatever hardware the supply chain delivers. This talk covers how vGPU interception enforces memory and compute limits, how the scheduler allocates heterogeneous devices, how DRA fits (and where it falls short), and production patterns from cloud providers running 10,000+ GPUs. No vendor pitch. A concrete blueprint for turning a GPU box into a GPU cloud.