| name | generate-openvino-stress |
| description | Generate sustained AI inference load on CPU, GPU, or NPU using OpenVINO benchmark_app in a container (K3s pod or Docker). Produces real neural-network inference stress for power/thermal profiling — unlike stress-ng synthetic load. Auto-detects whether K3s or Docker is available; supports both runtimes. Pod specs are generated inline (no external YAML files). Ideal for validating power profiles under realistic AI workloads, measuring inference throughput per watt, and thermal qualification with real compute patterns. |
Purpose
openvino_stress.sh generates a real AI inference load (via OpenVINO
benchmark_app) for power/thermal evaluation. Unlike stress_gen.sh which uses
synthetic stress-ng workloads, this script exercises the hardware with actual
neural-network inference — producing compute, memory, and accelerator access
patterns representative of production AI workloads.
Use it when you want to validate power/thermal behavior under realistic AI
inference load, or to compare power consumption between CPU, GPU, and NPU
inference for the same model.
Terminology
| Term | Meaning |
|---|
| benchmark_app | OpenVINO's inference throughput measurement tool. |
| OpenVINO | Intel's AI inference toolkit optimized for Intel hardware. |
| device | Target accelerator: CPU, GPU (Intel Xe via SR-IOV VF), or NPU. |
| runtime | Container orchestrator: K3s (kubectl) or Docker. Auto-detected. |
| niter | Number of inference iterations; 0 = run for a fixed duration instead. |
| duration | Time in seconds to run when niter=0; the benchmark loops until time expires. |
| nthreads | CPU threads dedicated to inference; GPU default is 2 (minimal CPU for offload). |
