You need one physical GPU split into isolated instances for several small jobs. Which feature does this?
Why D is the answer
Multi-Instance GPU (MIG)
MIG partitions supported GPUs into isolated instances with their own resources.
In detail
- MIG splits one GPU into isolated hardware instances.
- Only supported GPUs offer MIG.
Use case
An illustrative scenario, not a real incident.
Seven small inference jobs each use a fraction of a GPU. MIG gives each one its own isolated slice.
See it in different ways
flowchart TD
A{"What do you need?"} -->|"Hardware partitions"| B["MIG"]
A -->|"GPU in virtual machines"| C["NVIDIA vGPU software"]mindmap
root((Operations))
Monitoring
nvidia-smi
DCGM
Scheduling
Slurm
Kubernetes
Provisioning
Base Command Manager
Sharing
MIG
vGPUmindmap
root(("Physical GPU"))
n1["Instances"]
n2["Job A"]
n3["Job B"]
n4["Job C"]Key point: MIG partitions one GPU.
Why the other options are wrong
- A. InfiniBand partitioning
- That separates network tenants.
- B. Remote Direct Memory Access
- RDMA moves data between servers.
- C. NVLink Switch System
- That links GPUs together.