Several small workloads need isolated portions of a supported GPU's compute and memory. Which feature is relevant?

1. A mutable image tag
2. Multi-Instance GPU
3. A larger Kubernetes namespace name
4. CPU affinity alone