Skip to content

Frigate Host Benchmark: caneast-site1-node4 vs caneast-site1-node5

Status: Tier 3 Operational Reference Last Updated: 2026-05-03 WI: WI-402 Decision: caneast-site1-node4 selected. Phase 2 benchmark deliberately skipped.

Summary

Phase 1 hardware inventory collected 2026-05-03. The RAM gap was decisive; Phase 2 (OpenVino workload benchmarks) was skipped because node4 was already running Frigate successfully and the hardware delta left no realistic path to selecting node5.

Phase 1: Hardware Inventory

Property caneast-site1-node4 caneast-site1-node5
Role k3s worker + control-plane k3s worker
CPU Intel Core i9-9900K (8c/16t, 3.6-5.0 GHz) Intel Core i5-6500 (4c/4t, 3.2-3.6 GHz)
RAM total 30.7 GiB 7.1 GiB
RAM available ~29 GiB ~3-4 GiB (k8sgpt-ollama resident)
GPU Intel UHD 630 (Gen9.5, Coffee Lake, 9th gen) Intel HD 530 (Gen9, Skylake, 6th gen)
GPU EUs 24 24
GPU gen advantage Yes -- Gen9.5 has better INT8/throughput in OpenVino Baseline
k8s CPU allocatable ~15.5 cores ~3.9 cores
k8s memory allocatable ~29.5 GiB ~6.5 GiB
GPU resource gpu.intel.com/i915 (1 unit, allocated to Frigate) gpu.intel.com/i915 (1 unit, free)
Frigate currently running Yes (0.14.1, healthy) No

Decision: caneast-site1-node4

Selected. Reasoning:

  1. RAM gap is a dealbreaker. Frigate baseline is ~2 GiB resident. node5 has ~3-4 GiB available after k8sgpt-ollama. Future camera additions or event recording spikes would push node5 into OOM territory with no headroom. node4 has ~29 GiB free.

  2. GPU generation confirms the call. UHD 630 (Gen9.5) has demonstrably better INT8 throughput in OpenVino than HD 530 (Gen9). Both have identical EU counts (24) so the advantage is not count-based -- it is microarchitectural. This matters for detector frames/s.

  3. Proven in production. Frigate 0.14.1 was deployed to node4 as part of WI-394/WI-396 and has been running healthy. The intel-gpu-plugin DaemonSet has already allocated gpu.intel.com/i915 to the Frigate pod without issues. No unknown risk surface to validate.

  4. CP concern overridden by RAM math. node4 runs as both control-plane and worker (k3s single-node CP pattern). This is a valid concern for high-CPU workloads. However, Frigate's GPU-accelerated detector offloads the decode/detect path to the iGPU, keeping CPU utilization low. With 15.5 allocatable cores, CP overhead does not materially compete.

Why Phase 2 Was Skipped

Phase 2 was scoped to run OpenVino inference benchmarks (detector FPS, CPU% under load, memory pressure) side-by-side on both nodes. It was deliberately skipped for two reasons:

  • The Phase 1 hardware delta was unambiguous. There was no scenario where node5 benchmark numbers could override the RAM gap and the already-proven production deployment on node4.
  • Running Frigate benchmarks on node4 while it serves production traffic would introduce unnecessary disruption risk.

Phase 2 benchmarks remain available as a future data-gathering exercise if node capacity planning requires revisiting (e.g., multi-camera scale-out scenario).

Cluster State at Decision Time (2026-05-03)

Node         Role      GPU (i915)  CPU Alloc  RAM Alloc
caneast-site1-node4    CP+Worker 1/1 used    15.5c      29.5 GiB
caneast-site1-node5    Worker    0/1 free    3.9c       6.5 GiB

Frigate 0.14.1 running: archon-vision namespace, caneast-site1-node4. Intel GPU Plugin DaemonSet: kube-system, nodes caneast-site1-node4 + caneast-site1-node5.