Manage GPU Nodes
GPU node management is the same surface as regular node management — same Nodes page, same actions menu, same lifecycle, same network and security controls. This page only documents the GPU-specific differences. For everything else, follow Manage Nodes.
What's GPU-Specific
| Area | GPU-specific difference | Where the regular-node reference lives |
|---|---|---|
| Details | The node details page includes the GPU card, GPU count, GPU memory, and CUDA version in addition to the standard CPU/RAM/disk/IOPS fields. | Manage Nodes |
| Monitoring | The Monitoring tab adds per-card graphs for GPU utilization, GPU memory (used/free), GPU power, and GPU temperature. Per-card series are labeled card_1, card_2, card_3 on multi-card plans. | Node Monitoring and Alerts |
| State transitions | Same states as regular nodes (Creating, Running, Powered off, Stopped, Saving, Reinstalling, etc.). Transitions are slower because the NVIDIA driver and CUDA toolkit re-initialize on each boot. | Node states |
| Save Image | Captures the entire root disk, including the NVIDIA driver, CUDA toolkit, container caches, and framework installs. GPU saved images can reach hundreds of GB and incur storage charges for the full size while retained. | Save Image |
| Plan Upgrade | Same family upgrades (1× → 2× → 4× → 8×, higher vCPU/RAM SKU) are allowed. Cross-card-type upgrades (for example A100 → H100, L4 → L40S) are not supported from the portal — migrate via a saved image. | Upgrade Plan |
| Reinstall | Wipes the root disk, including the NVIDIA driver and CUDA toolkit. If you do not have a saved image, plan for a full driver-and-framework install path on first boot. | Reinstall |
| Billing | GPU plan SKUs use the GPU-, GDC-, or GDC3- prefixes. Committed GPU plans cannot be downgraded for the duration of the committed period. | Committed Nodes |
Everything not in this table — Start, Power Off, Reboot, Lock/Unlock, Move to Another Project, Convert to Committed, Add/Detach Volume, Accidental Protection, Recovery Mode, Security Compliance, Delete, Bulk Actions, security groups, BitNinja, VPC, IPv6 — is the same as a regular node. Use Manage Nodes and Node Actions as the canonical reference.
Snapshots are not available for GPU nodes. Use Save Image to capture a re-launchable template of the root disk instead.
Available actions depend on the node state, plan family, OS, project permissions, region, attached services, and billing status. The portal disables actions that are not valid for the selected node. See Action Availability.
Open the Manage Nodes Page
- Log in to MyAccount.
- Go to Compute > Nodes. CPU and GPU nodes are listed together; filter or sort by plan to narrow to GPU nodes.
- Select a node name to open its details page, or use the Actions menu from the node list.
Related Resources
| Resource | Use it for |
|---|---|
| Manage Nodes | Canonical management reference for all the non-GPU behavior. |
| Node Actions | Full action reference (Start, Stop, Save Image, and so on). |
| Action Availability | Why an action might be hidden, disabled, or rejected. |
| Node Monitoring and Alerts | Alerting, delivery, and the non-GPU monitoring graphs. |
| Committed Nodes | Committed plan rules. |
| Create a GPU node | Launch a GPU node. |
| Choose a GPU card | Compare cards before launch. |
| Troubleshoot GPU Nodes | NVIDIA driver, CUDA, GPU-not-detected, OOM, throttling fixes. |