Feature Overview:

Enable seamless GPU observability in OpenShift by integrating NVIDIA DCGM (Data Center GPU Manager) metrics into the OpenShift Monitoring stack. This feature will provide one-click solution to collect, export, and visualize GPU telemetry using prebuilt dashboards.

Use Case:

OpenShift administrators running AI/ML workloads on GPU nodes often need deep insights into GPU utilization, temperature, memory, and clock events. Manually instrumenting and collecting these metrics is complex and error-prone. This feature simplifies GPU observability with minimal setup effort.

Key Metrics Collected:

• GPU Utilization (%)

• SM Utilization (%)

• GPU Temperature (°C)

• Framebuffer Memory Utilization (%)

• Clock Throttle Events (avg by model)

• Streaming Multiprocessor (SM) Clock Speed

• Power Consumption (W)

• I/O and memory throughput