Can I share a dedicated server GPU using MIG or passthrough?
GPU sharing and assignment depend on the exact hardware and software stack. Whole-device passthrough assigns a device to a guest; NVIDIA MIG partitions supported GPUs into defined resource instances. Neither feature should be assumed available on every GPU or interchangeable with ordinary container sharing.
Before you start
- Exact GPU model, host OS, hypervisor and intended application requirements.
Instructions
1. Confirm the supported matrix
Check the vendor’s specific GPU list, driver requirements, operating systems and hypervisors. A GPU being from a recent architecture does not establish support for every partitioning feature. Confirm any virtual-GPU licence and provider support requirements before selecting the design.
2. Match the resource boundary
Compare required memory, compute resources, device features and workload isolation with the supported profiles. Consider how multiple processes share a whole GPU versus the resources assigned to a MIG instance. A guest or container must be configured to receive the intended supported device or instance.
3. Plan configuration as maintenance
Review the model-specific procedure for enabling partitioning, creating instances or assigning passthrough. Such changes may require draining jobs, resetting the device or restarting components. Rehearse application compatibility, allocation limits and recovery behaviour before changing a busy server.
Check the result
- The proposed sharing mode appears in the supported hardware/software matrix.
- Each workload has tested access to the intended resources and required features.
Troubleshooting
- A GPU supports CUDA but cannot enable MIG
- Check the exact supported-product and deployment lists. CUDA capability alone does not establish that the model, driver and operating mode support the requested MIG configuration.
Official references
Use the documentation for your installed version when a flag, service name or platform feature differs.