TIR Private Cluster FAQ

For step-by-step fixes, see Troubleshooting TIR Private Cluster.

Clusters and projects

What is the difference between a private cluster and an on-demand instance?

A private cluster is a dedicated pool of GPU nodes that your account owns and your projects share, billed per node at a fixed rate however much you use it. The services you run on it are included in that rate. A single on-demand instance is drawn from E2E’s available stock and billed by the hour only while it runs. Choose a private cluster for long-running or shared work, and an on-demand instance for something short-term or variable. See TIR Private Cluster.

Who can create or delete a private cluster?

Only the Owner or an Admin. They also change a cluster’s configuration and allocate nodes to projects. Team Leads, Project Leads, Project Managers, and IAM users can view the cluster and take nodes back from a project they can update; the cluster is hidden from Team Members and Members. See Nodes, projects, and access.

Which GPUs can a private cluster use?

H200, H100, A100, L40S, and L4, the same range TIR offers for its GPU nodes, each on hourly or committed plans. See Create a private cluster.

Can I share a private cluster across multiple projects?

Yes. You can split a private cluster’s nodes across several of the account’s projects at once, and each project gets its own share. Every node can run any service type, so allocated nodes in different projects can run different kinds of service at the same time. See Allocate nodes to a project.

Can I reduce the node count of a cluster?

No. Update Configuration only adds nodes to a cluster. To run with fewer nodes, delete the cluster and create a new one with the count you want. See Update a private cluster and Delete a private cluster.

Can I delete a single node and keep paying for the rest of the cluster separately?

No. A cluster’s nodes are removed together, when you delete the cluster. To run with fewer nodes, delete the whole cluster and create a new one with the count you want. See Delete a private cluster.

Why can I not delete my private cluster?

A service is still on it. TIR refuses the delete while any instance, model endpoint, training cluster, vector database, or Slurm cluster is on the cluster, even a stopped one, and the message names each one. Delete them, then delete the cluster. The nodes can stay allocated to their projects; they are released with the cluster. See Delete a private cluster.

Does deleting a cluster delete my file systems and datasets?

No. A Shared File System, Parallel File System, or dataset attached to a service on the cluster stays, and keeps its own billing, until you delete it where you manage it. See Billing options.

What happens to my private cluster if my account is deprovisioned?

Every private cluster you still have is deleted with the account, even one that still has services on it. Back up anything you need before the account is deprovisioned.

Nodes and allocation

Can I move a node with a running service directly to a different project?

No. A move takes nodes only from the cluster’s free nodes. Stop or delete the service, deallocate the now-free node from its project, then allocate it to the other project. A service belongs to its project, so create it again in the new one. See Allocate nodes to a project.

Can a project member take nodes back from their project?

Yes, from their own project’s share, if they have update access to the project. Only the Owner or an Admin can add nodes to a project or move nodes between projects. See Allocate nodes to projects.

Why can I not take the last node off a project?

The project still has a stopped instance, model endpoint, or vector database on the cluster. TIR lists each one by name and status; delete them, then remove the node. See Removing the last node from a project is refused.

Does changing a node group move workloads that already use it?

No. An instance or model endpoint already running under a group’s rule stays where it is when you change the group. To move it onto the updated group, remove the rule from it, update the group, then apply the rule again. See Use node affinity.

Billing

Can I stop billing for individual nodes without deleting the cluster?

No. An hourly cluster is billed for every node in it, every hour it exists, whether a node is idle or running a workload. To stop paying for the cluster’s nodes, delete the whole cluster. See Billing options.

Can I add more nodes to an existing cluster?

Yes, from Update Configuration. The new nodes are billed hourly from when you add them, even on a committed cluster, because the term covers only the cluster’s original nodes. See Update a private cluster.

Is allocating or moving nodes charged?

Allocating, taking back, and moving nodes within the cluster is free, and so are security groups, VPC IPs, and IAM roles and policies used with it. See Billing options.

What happens to nodes I added to a committed cluster when the term renews?

They stay on hourly billing. Auto Renewal renews only the nodes in the original commitment, so nodes you added later keep billing by the hour. See Billing options.

Can I switch a committed cluster to hourly billing before the term ends?

No. The cluster stays on committed billing until the term ends. Then the end-of-term choice you made at creation takes effect: Auto Renewal, Convert to Hourly Plan, or Auto Delete.

Does commitment renewal happen on its own, or does E2E confirm it with me first?

It happens automatically. TIR emails you after the renewal, or after the switch to hourly billing if your balance was too low to cover it.

Is there a refund if I delete a committed cluster early?

No. You pay for the whole term upfront, and the time left stays charged when you delete the cluster early.

How long does it take to create a private cluster?

A few minutes, depending on the node count and GPU type you choose. When it is created, TIR opens Allocate Nodes to Projects so that you can give its nodes to a project straight away. If a cluster stays in a creating state for more than about 10 minutes, see Troubleshooting TIR Private Cluster.

MIG

Can I enable MIG on any GPU instance, or only inside a private cluster?

Only inside a private cluster. To run a service on part of a GPU, launch it on a MIG-enabled private cluster. See GPU partitioning with MIG.

Can I turn MIG on or off after creating the cluster?

Not in the portal. You set MIG Enable for the whole cluster when you create it, and the portal keeps that setting for as long as the cluster exists. See Create a private cluster.

Which MIG instance sizes can I use, and how many fit on one GPU?

On an H100, the sizes run from 1g.10gb to the whole GPU, 7g.80gb; on an H200, from 1g.18gb to 7g.141gb. Instances share a GPU in fixed layouts of up to seven, such as six 1g.10gb and one 1g.20gb on an H100, and you choose the size when you launch a service. See GPU partitioning with MIG.