Last updated · Fusionative
Short answer
GPU multi-tenancy on Kubernetes means letting several teams or customers share GPU capacity safely. It combines a GPU sharing method (whole GPUs, NVIDIA MIG partitions, time-slicing or vGPU) with a tenant isolation model (namespaces, virtual clusters or dedicated clusters), plus quotas and metering per tenant.
Ways to share a GPU
Whole-GPU allocation gives each workload full GPUs and the strongest isolation. NVIDIA Multi-Instance GPU (MIG) partitions supported data-center GPUs into isolated instances, up to seven, each with its own memory and compute. Time-slicing lets several pods take turns on one GPU without memory isolation, which suits development but not untrusted tenants. NVIDIA vGPU shares GPUs between virtual machines with a licensed driver.
Ways to isolate tenants
- Namespaces with RBAC, network policies and resource quotas: cheapest, weakest isolation.
- Virtual clusters: each tenant gets its own Kubernetes API on shared nodes.
- Dedicated clusters or VMs per tenant: strongest isolation, highest overhead.
Quotas, fairness and metering
Sharing only works if usage is bounded and visible. Kubernetes ResourceQuotas cap GPU requests per namespace; queueing systems such as Kueue add fair sharing and borrowing. Metering per tenant, by GPU-hour or GPU partition, is what turns sharing into showback, chargeback or billing.
Where FusioNative fits
FusioNative provides the tenancy layer above Kubernetes: tenants and projects, quotas for GPUs by type, including slices and fractions, enforced as Kubernetes quotas, VMs that use passthrough, time-sliced, vGPU or MIG devices, and metering of GPU-hours by GPU model for invoices.
| Method | Isolation | Best for | Trade-off |
|---|---|---|---|
| Whole GPU | Strong | Training, large inference, untrusted tenants | Low utilization for small jobs |
| MIG partitions | Strong (hardware) | Many small inference workloads, multi-tenant | Supported GPUs only, fixed profiles |
| Time-slicing | Weak (no memory isolation) | Development, trusted teams | Noisy neighbors, no hard limits |
| vGPU | Medium to strong | Virtual machines needing a GPU share | Licensed software, hypervisor required |