You are setting up a monitoring pipeline for a Kubernetes cluster running GPU-intensive AI workloads. You need to collect telemetry data from NVIDIA GPUs and feed it into a Prometheus database. Which software component must you deploy to export these specific metrics?
Select an answer to reveal the explanation.
Short Explanation and Infographic
If you want to keep your GPUs healthy and running hot, you need real-time data on temperature, memory usage, and compute load. NVIDIA has a tool called DCGM (Data Center GPU Manager) that does the heavy lifting, but Prometheus doesn't speak DCGM out of the box. That's where the NVIDIA DCGM Exporter comes in. It acts as a translator, grabbing the metrics from DCGM and serving them up in a format Prometheus loves. Trust me, you'll want this set up so you can build those beautiful Grafana dashboards!
Full explanation below image
Full Explanation
Monitoring GPU performance and health is critical in large-scale AI infrastructure. The primary tool for managing and monitoring NVIDIA GPUs in a cluster is DCGM (Data Center GPU Manager). To integrate DCGM telemetry with cloud-native monitoring pipelines, the NVIDIA DCGM Exporter is used. The Exporter acts as a bridge, querying DCGM for low-level GPU metrics (such as Tensor Core utilization, memory usage, temperature, and power consumption) and exposing them on HTTP endpoints in Prometheus-compatible exposition format. Prometheus can then scrape this endpoint periodically, allowing administrators to visualize GPU health and set up alerts using tools like Grafana. - Option A is incorrect because nvidia-smi is a command-line utility designed for single-node status checks. It is not designed to export cluster-wide metrics to time-series databases like Prometheus. - Option B is incorrect because the NVIDIA Container Toolkit allows containers to access host GPU resources; it does not export monitoring telemetry. - Option C is correct as it is the dedicated component designed to translate DCGM metrics into Prometheus format. - Option D is incorrect because the NVIDIA GPU Operator is an orchestration tool that automates the deployment of all GPU-related software components (including the driver, Container Toolkit, and DCGM Exporter) in Kubernetes, but it is not the component that directly collects and exports the telemetry metrics itself.