The 2026 Hardware Shift: For years, the default advice for building an on-premises Frigate NVR surveillance system was to track down an elusive $60 Google Coral USB accelerator on eBay. In 2026, the Google Coral TPU is officially legacy hardware. Modern Intel integrated GPUs (iGPUs) running OpenVINO and modern M.2 coprocessors (like the Hailo-8) vastly outperform the Coral in both frames-per-second throughput and model flexibility (YOLOv8/v9/v10). Here is the definitive hardware sizing matrix, power benchmarks, and network architecture for running zero-cloud, 24/7 computer vision at home.

1. The Two-Stage Video Pipeline: Decode vs. Detect

Sizing hardware for Frigate requires understanding that video processing is split into two completely separate physical workloads:

  1. Hardware Video Decoding (H.264 / H.265): Taking raw compressed RTSP streams from your IP cameras and decompressing them into raw video frames in memory. This is performed by your CPU’s hardware video decoder (e.g., Intel Quick Sync Video or VAAPI).
  2. Object Detection (Inference): Running a convolutional neural network or YOLO vision model across isolated bounding boxes to classify objects (person, car, dog, package). This is performed by an AI accelerator (Intel iGPU via OpenVINO, Hailo-8, or Coral).

The Golden Rule: Never Run Detection on 4K Main Streams

A 4K (3840×2160) stream at 30 fps delivers over 248 million pixels per second. Running object detection on full-resolution 4K streams will instantly peg an 8-core CPU to 100% and overheat your server.

The Standard Configuration: Configure dual streams on your cameras. Point Frigate’s record engine to the 4K high-bitrate main-stream, and point the detect engine to a dedicated sub-stream at 720p (1280×720) or 640×360 running at 5 frames per second. A 5 fps sub-stream provides 99.8% detection accuracy while consuming 90% less compute.

2. AI Accelerators Showdown: OpenVINO, Hailo-8, Coral, and TensorRT

How do modern edge AI processors stack up for residential Frigate deployments?

Accelerator Inference Engine TOPS (INT8) Inference Speed Model Support Hardware Cost
Intel iGPU (UHD 730/770 / Iris Xe)
11th–14th Gen & N100/N200
Intel OpenVINO Integrated (Shader execution) 6 ms – 12 ms YOLOv8, YOLOv9, MobileNet $0 (Built into CPU)
Hailo-8 / Hailo-8L
M.2 2280 / 2242 PCIe
HailoRT / OpenVINO wrapper 13 to 26 TOPS 2 ms – 4 ms YOLOv8x, YOLOv9c, Custom CNNs $70 – $130
Google Coral TPU
USB / M.2 (Legacy)
libedgetpu 4 TOPS 8 ms – 14 ms Obsolete MobileNet-SSD only $60–$120 (Inflated aftermarket)
NVIDIA Discrete GPU
GTX 1650 / RTX 3050
NVIDIA TensorRT 30 to 100+ TOPS 1 ms – 3 ms All modern architectures $150–$300 + 40W–75W idle penalty
Rockchip RK3588
Orange Pi 5 Plus / Firefly
RKNN Toolkit 6 TOPS (Tri-core) 5 ms – 9 ms YOLOv5, YOLOv8 $110 – $150 (Complete board)

3. Hardware Sizing Tiers: Matching Cameras to Hardware

Tier 1: Small Home (1 to 5 Cameras) — The Intel N100 Mini PC

Recommended Machine: Beelink EQ13 or Minisforum UN100D (Intel N100, 4 cores, 16GB DDR5, 500GB NVMe, Dual 2.5GbE NICs).

  • Decode: Intel Quick Sync handles up to 6 camera sub-streams at <15% CPU load.
  • Detect: OpenVINO running on the Intel UHD Graphics iGPU delivers 8–10 ms inference on YOLOv8n.
  • Idle Power: Incredible 6W to 9W whole-system wall draw.
  • Total Hardware Cost: ~$150 – $180 complete.

Tier 2: Mid-to-Large Estate (6 to 14 Cameras) — Intel Core i5 Desktop / SFF

Recommended Machine: Refurbished HP ProDesk / Dell OptiPlex 7000 SFF (Intel Core i5-12500 or i5-13500, 32GB RAM, Intel UHD 770).

  • Decode: Dual Quick Sync media engines on the UHD 770 decode up to 16 camera streams with zero CPU bottleneck.
  • Detect: OpenVINO iGPU inference operates at 5–7 ms. Has ample headroom to co-host Home Assistant Core, Zigbee2MQTT, and Plex/Jellyfin.
  • Idle Power: 18W to 24W wall draw.
  • Total Hardware Cost: ~$300 – $400.

Tier 3: Enterprise / Heavy AI (15 to 24+ Cameras) — Intel Core Ultra + Hailo-8

Recommended Machine: Intel Core Ultra 5 125H / Core Ultra 7 SFF with dedicated M.2 Hailo-8 (26 TOPS) coprocessor.

  • Decode: Intel Arc-class Xe-LPG graphics engine handles high-throughput AV1, HEVC, and H.264 streams.
  • Detect: Dedicated Hailo-8 handles 100+ frames per second across YOLOv9c models with sub-3 ms latency, freeing the CPU completely for running on-premise LLMs (Ollama) and local speech synthesis.
  • Idle Power: 22W to 32W wall draw.

4. Camera Security: Dual-NIC Architecture & VLAN Isolation

IP cameras from vendors like Reolink, Amcrest, and Dahua are notorious for calling home to Chinese cloud servers and probing local subnet peers. Never place IP cameras on your primary home network.

The gold-standard local-first architecture uses a mini-PC with Dual Physical Network Interfaces (NICs):

  • NIC 1 (ETH0 – Subnet 192.168.1.x): Connects to your trusted home LAN / Home Assistant server. Has outbound internet access for remote viewing via WireGuard / Tailscale VPN.
  • NIC 2 (ETH1 – Subnet 10.10.20.x): Connects to a dedicated PoE switch powering the IP cameras. This interface has no default gateway and zero internet access.

Cameras are physically incapable of reaching the internet or pinging external cloud NTP servers. Frigate acts as the sole cryptographic bridge, ingesting the local RTSP streams on ETH1 and serving encrypted WebRTC streams to trusted devices on ETH0 via Privacy Cameras standards.

5. Storage Engineering: NVMe SSDs vs. Surveillance HDDs

One of the fastest ways to kill a local NVR is writing 24/7 continuous video footage onto a consumer TLC NVMe SSD. Consumer SSDs have finite write endurance (TBW) and will burn out their flash cells within 12 to 18 months under constant camera recording.

  • System & Database Drive: Use a high-end 500GB NVMe SSD (Samsung 980 Pro or Crucial T500) exclusively for the host OS, Docker containers, Frigate event clips, and the SQLite/MariaDB database.
  • 24/7 Continuous Recording Drive: Use a dedicated Surveillance-Rated Mechanical HDD (Western Digital Purple or Seagate SkyHawk) connected via SATA or USB 3.2 UASP enclosure. These drives feature specialized firmware designed for continuous sequential writes and thermal resilience.

Frequently Asked Questions

Why is Google Coral no longer recommended for new Frigate builds?

The Google Coral Edge TPU is limited to 4 TOPS and relies on an outdated TensorFlow Lite compiler that only runs legacy MobileNet-SSD models. Modern Intel integrated GPUs running OpenVINO outperform the Coral at zero additional hardware cost, while newer coprocessors like the Hailo-8 deliver 26 TOPS and support modern YOLOv8/v9 vision architectures.

Can an Intel N100 mini-PC handle 8 cameras on Frigate?

Yes, provided you configure dedicated 720p 5fps detection sub-streams on all 8 cameras and enable Intel Quick Sync video decoding with the OpenVINO detector. If you attempt to run detection on uncompressed 4K main streams, the N100 will bottleneck.

Do I need a separate graphics card for Frigate NVR?

No. A dedicated discrete graphics card (like an NVIDIA GPU) is unnecessary for residential systems and introduces a severe 40W+ continuous idle power penalty. Modern Intel CPUs with integrated Quick Sync graphics handle video decoding and object detection effortlessly within a 10W–20W power envelope.

// LOCAL-FIRST HOME AUTOMATION & PROTOCOLS

// The Local Mesh

Bi-weekly architectures for local-first smart homes, Home Assistant YAML patterns, Matter/Thread co-existence, and private NVR pipelines.