GPU Detail
The GPU Detail page lets you monitor the GPU usage of a specific server with time-series charts and view the processes occupying the GPU in real time. Physical GPU and MIG (Multi-Instance GPU) instance information are available on a single screen, allowing you to quickly identify which workloads are consuming which GPU resources and by how much.
Supported Agent Version
Supported on Linux only. Agent version 2.9.16 or later is recommended.
Main Screen
In the options bar at the top of the screen, you can select the time range and server to query. The selected server's inline information (IP, OS Type, CPU Cores, Memory Total, OS Version) and a Feature Activation Guide button are displayed together. Below the options bar, you can filter the GPUs to display by selecting them.
Usage for each item is as follows:
- Time Range: Select the time period to query. Use the left and right arrow buttons to navigate between periods.
- Server Selection: Select the server to monitor from the dropdown. The hostname and IP are displayed together.
- GPU Selection: Filter the GPUs to display. Click the selection box to open the GPU selection modal. The default is
* (Select All), which selects all Physical GPUs and MIG instances connected to the server. - Feature Activation Guide: Click to open a panel with information on the GPU monitoring supported environment and the list of collectible metrics.
GPU indexes are displayed in the following format across chart legends and the process table:
- Physical GPU: Integers such as
0,1,4 - MIG Instance: Slash notation such as
7/1/0,6/4/0,5/8/0(Physical/GI/CI)
GPU Metric Charts
The left area of the screen displays the GPU usage of the selected server as five time-series chart cards. These consist of 2 line charts (both Physical GPU + MIG instances) and 3 stack/line toggle charts (Physical GPU only).
You can switch to line mode using the toggle button in the upper right of each card.
| Chart | Type | Description |
|---|---|---|
| GPU Utilization (%) | Line | GPU core utilization |
| FB Memory Used (%) | Line | GPU frame buffer memory utilization |
| Power Usage | Stack/Line | GPU power consumption |
| PCIe TX+RX | Stack/Line | PCIe transmit/receive bandwidth |
| NVLink TX+RX | Stack/Line | NVLink transmit/receive bandwidth |
Using the icons in the upper right of each chart card, you can download the data as a CSV file or save the chart as an image. Clicking the expand icon opens the Chart Detail Drawer for that metric category.
For detailed definitions of each metric, refer to the GPU Metrics document.
Chart Detail (Drawer)
Clicking the expand icon in the upper right of each chart card opens a drawer on the right side where you can view the detailed metrics for that metric category. Within the drawer, you can change the time range and GPU selection independently from the main screen.
In the Interpretation Guide section at the bottom of the drawer, you can check the meaning and monitoring points for each metric.
| Drawer | Included Metrics |
|---|---|
| GPU Utilization Detail | GPU core utilization, SM Active/Occupancy, FP·Tensor·DRAM activity, etc. |
| GPU Memory Detail | FB Total/Used/Free/Reserved, FB Used (%), memory copy utilization, etc. |
| Power/Temperature Detail | Power Limit/Usage, power violation rate, GPU temperature, thermal violation rate, etc. |
| PCIe Traffic Detail | PCIe TX+RX, PCIe TX, PCIe RX, NVLink TX+RX, etc. |
| NVLink Traffic Detail | NVLink TX+RX, NVLink TX, NVLink RX, PCIe TX+RX, etc. |
GPU Process Table
The GPU Process table in the upper right of the screen displays the list of processes occupying the GPU on the selected server. You can check the GPU, CPU, and memory usage trends of each process at a glance, and filter by process name using the search box in the upper right of the table.
Each metric cell displays a time-series mini chart along with the maximum value (Max) for the query period.
| Column | Description |
|---|---|
| Process Info | Process name, execution command, PID xxxx · account (3 lines) |
| GPU Utilization (%) | GPU Utilization of the entire GPU device to which the process is bound. This is not a directly measured value per process but shows the overall utilization of the connected GPU, so multiple processes using the same GPU may show identical or similar values. Mini chart + Max N% |
| FB Used (MiB) | GPU frame buffer memory size occupied by the process. Mini chart + Max N MiB |
| CPU (%) | CPU utilization of the process. Mini chart + Max N% |
| RSS (byte) | Host system memory usage of the process (Resident Set Size). Mini chart + Max N GiB |
| Process IO Total (bytes/s) | I/O bandwidth of the process. Mini chart + Max N KiB/s. Displays - if no data is collected |
| GPU Index | GPU index occupied by the process. Physical is an integer (0, 1), MIG uses slash notation (5/8/0) |
VLLM::Worker_TP0, VLLM::EngineCore, python3 app.py, etc. are the names of processes actually running in the user environment. All processes occupying the GPU are displayed in the table.
Server Information · GPU Inventory Panel
In the lower right panel, select the Server Information or GPU Inventory tab to view detailed information about the server currently being queried. You can adjust the ratio between the GPU process table and this panel using the drag handle between them.
Server Information
| Item | Description |
|---|---|
| Host name (oid) | Hostname and OID |
| OS | Operating system type and version |
| GPU model | GPU model name (e.g., A100-40GB) |
| GPU count | Number of GPUs connected to the server (in the format Physical N / MIG M) |
| Server memory | Total server memory |
| Agent | Installed agent version |
GPU Inventory
You can view the list of Physical GPUs and MIG instances connected to the server with the following columns. Click the Detail View button on the right side of the tab to open a drawer showing all columns.
Only the inventory at the current point in time can be queried. To query previous points in time, use the GPU Inventory menu.
| Column | Description |
|---|---|
| GPU Index | Physical is an integer (0, 1), MIG uses slash notation (5/7/0, 5/8/0) |
| GPU Type | Physical or MIG |
| Model Name | Official GPU model name (e.g., NVIDIA A100-SXM4-40GB) |
| Model Family | GPU model family (e.g., A100) |
| GPU Memory Size (MiB) | GPU memory size. Physical shows total memory, MIG shows the partitioned fragment size |
For the inventory menu itself and how to use it, refer to the GPU Inventory page.