General GPU vs. dedicated GPU

• General GPU container: Select an available GPU from the dashboard. Availability can change in real time.
• Dedicated GPU reservation: A contract-based resource reserved for a specific organization during a defined time window.
• MIG partitioned resource: A supported physical GPU is split into multiple isolated GPU instances.

What is MIG?

MIG, or Multi-Instance GPU, is an NVIDIA feature that partitions GPU memory and compute resources into multiple instances on supported GPUs. StreamGPU may describe these as partitioned GPU instances or MIG instances.

Not every GPU supports MIG, and the available partition profiles depend on the GPU model and operational state.

When to use it

• Your organization needs stable GPU capacity for a known period.
• Multiple members need smaller isolated GPU units.
• The workload is predictable, such as education, labs, or inference.
• Contracted capacity is more important than shared pool availability.

Request and setup flow

1. Share the organization name, desired GPU, time window, and expected usage with StreamGPU sales or support.
2. StreamGPU checks GPU model, availability, MIG support, and estimated cost.
3. After approval or contract confirmation, an organization resource reservation is created.
4. If MIG is required, StreamGPU runs safety checks before applying the configuration.
5. After setup, organization containers can use the reserved resource.

Operational cautions

• MIG configuration changes may affect workloads running on the target GPU.
• Coordinate job shutdown, data backup, and the change window before reconfiguration.
• Dedicated reservations have start and end times; usage conditions may change after the period ends.
• Insufficient organization balance or contract limits can restrict new creation or renewal.

Billing

Dedicated GPU and MIG instance billing may differ from normal pay-as-you-go container billing. Depending on the contract, prepaid reservations, organization balance, or separate invoicing may apply.

Use the quotation or contract document as the source of truth for exact rates and time windows.

Information to include in requests

• Organization name or customer code
• Required GPU model or VRAM/performance needs
• Required period and expected usage window
• If MIG is needed, desired instance count and approximate memory requirement
• Container image, framework, and expected workload