Skip to content

GPU Dashboard

Use the GPU Resource Dashboard once NVIDIA GPU Operator is deployed. NVIDIA GPU Operator deployment automates the management of all NVIDIA software components, which are required to provision GPUs within Kubernetes. The GPU Resource Dashboard is powered by the dcgmExporter deployed as part of the GPU Operator, and it gives deeper visibility into the resources of your GPU cores.

How it Works

The Rafay managed Prometheus automatically scrapes the GPU metrics (if available) and aggregates them in a multi-tenant, time series database on the controller. The diagram below describes the process at a high level. 

sequenceDiagram
    autonumber
    participant user as Data Scientist
    participant rafay as Rafay SaaS

    box Kubernetes Cluster in Remote Datacenter 
    participant prom as Managed Prometheus 
    participant operator as GPU Operator
    participant gpu as Nvidia GPU
    end

    operator->>gpu: Retrieve GPU Metrics
    prom->>operator: Scrape GPU Metrics 
    prom->>rafay: Aggregate Metrics 
    user->>rafay: Access/View GPU Metrics

The GPU Dashboard contains the graph of current GPU SM Clocks, Current GPU Memory Clocks, GPU Utilization, GPU Memory Copy Utilization, GPU SM Clocks, GPU Memory Clocks, Framebuffer Memory Used, Framebuffer Memory Free, GPU Average Temperature, and GPU Power Tool

GPU Dashboard

Access a specific GPU core dashboard with the following options:

From Cluster Card

Click on GPUs count in the Cluster card of the project to view a list of all GPU cores of the cluster, the node they belong to and the pods using GPU cores

GPU Cores List

From the GPU core list, click node GPU core to view the GPU dashboard for that specific GPU core

GPU Core

From Node Dashboard

From the Node's dashboard, click GPU tab and select the GPU core in the drop down list to view the GPU dashboard for that specific GPU core.

Node GPU Core