GPU Dashboard

The GPU Dashboard provides a Kubernetes perspective to track the flow from GPU resource allocation to actual utilization and the per-device status on a map.

• Provides a funnel of stage-by-stage resource flow from Total to Effective so you can see actual utilization against allocation at a glance.

• Displays device-level resource distribution and status by cluster, node, and namespace on the GPU Map.

• Provides a toggle that filters waste resources only, so you can quickly identify GPUs that are allocated but not utilized.

Note

The GPU menu consists of the GPU Dashboard, GPU List, GPU Trend, and GPU Workload pages. This document covers the GPU Dashboard page, which provides resource utilization status and the device map.

Permissions & Requirements

• Supported environment: Kubernetes cluster project

• Agent version: Kubernetes Agent v1.8.7 or higher

• Requires Open Agent installation

Main Screen

A visualization dashboard that lets you identify GPU resource utilization status and per-device status within the cluster at a glance.

At the top of the screen are the LIVE time and the Show waste resources only toggle. Turning the toggle on displays only the devices classified as waste resources on the GPU Map.

Note

• The Show waste resources only toggle filters the display targets of the GPU Map and does not affect the aggregated values in the GPU Utilization Status funnel.

• Selecting the pause button in the time selector at the top of the screen lets you query data from a past time.

GPU Utilization Status

Provides a funnel chart of the flow from GPU resource allocation to actual utilization. The stages narrow in the order of Total, Allocated, Active, and Effective, and resources not utilized at each stage are separated and shown as Idle and Not Effective.

Stage	Description
Total	All physical devices
Allocated	Devices with a process or Pod allocated
Active	Devices with a 5-minute average GPU Utilization of 1% or higher
Idle	Devices with a 5-minute average GPU utilization below 1%
Effective	Devices with GPU utilization above 50%
Not Effective	Devices with GPU utilization of 50% or lower

In the example above, 8 of 8 Total are Allocated, split into 4 Active and 4 Idle, and the Active are classified into 2 Effective and 2 Not Effective.

Idle and Not Effective are devices that are allocated but not sufficiently utilized, serving as the basis for checking resource waste.

GPU Map

Displays the GPU devices collected at the query time on a map. Each device is placed as a box inside a node card, and the top of the card shows the device, namespace, service, and model counts for that node along with a GPU utilization summary.

Controls

Control	Option	Description
Group	Cluster / Node / Namespace / Service / Model	Criterion for grouping devices
Color by	Status / Utilization / VRAM(%) / Temp	Metric for the device box color
Util basis	Average / Minimum / Maximum	Statistic used for the utilization display
Group by physical device	On / Off	Group and display devices by physical device
Group name search	Text input	Filter display targets by group name

Node Cards and Device Boxes

Each node card displays the devices installed on that node as boxes. The summary at the top of the card shows the following information.

• Device n (Physical n): Device count and physical device count

• Namespace n / Service n / Model n: Namespace, service, and model counts associated with the node

• GPU Util (min / avg / max): Minimum, average, and maximum of the node GPU utilization

Device boxes are displayed differently by status.

Display	Meaning
MIG-labeled box	Device split into MIG instances
Colored box	Active device, color determined by the Color by criterion
Gray box	Idle device (5-minute average GPU utilization below 1%)

The legend distinguishes Active and Idle by color.

Hovering over a device box shows the device type (Physical/MIG), GPU Utilization, Status, Device, Pod, Node, Namespace, and Service information.

Note

• Changing the Color by option makes the color of the same device differ depending on the criterion metric (status, utilization, VRAM, temperature).

• A device split by MIG shows one physical device as multiple instances, so the device count and physical device count can differ.