Frigate NVR development

1. High-Level Product Architecture 📊 App Core & Hardware Interaction Pipeline This diagram illustrates the separation of concerns. Video decoding is isolated to Intel Silicon, while AI inference and semantic queries are mapped directly to the Nvidia Blackwell Tensor cores. [ ANY ONVIF / RTSP CAMERA ] │ │ (ONVIF Video & PTZ Control Streams) ▼ ┌────────────────────────────────────────────────────────────────────────┐ │ INTEL i5 CPU (BASE UBUNTU LINUX HOST) │ │ │ │ ┌───────────────────────────────┐ ┌──────────────────────────────┐ │ │ │ Intel UHD 730 QuickSync │ │ System RAM (32GB DDR5) │ │ │ │ (0% CPU Hardware Decoding) │ │ (1GB tmpfs Live Video RAM) │ │ │ └───────────────┬───────────────┘ └──────────────┬───────────────┘ │ └──────────────────┼──────────────────────────────────┼──────────────────┘ │ Video Frames │ Cached Event Segments ▼ ▼ ┌────────────────────────────────────────────────────────────────────────┐ │ DOCKER CONTAINER SYSTEM (FRIGATE ENGINE) │ │ │ │ ┌──────────────────────────────────────────────────────────────────┐ │ │ │ NVIDIA RTX 5060 Ti (16GB VRAM BLACKWELL) │ │ │ │ - TensorRT Object Tracking Pipeline │ │ │ │ - Local CLIP/Jina Embedding Processing Pipeline │ │ │ └───────────────────────────────┬──────────────────────────────────┘ │ └──────────────────────────────────┼─────────────────────────────────────┘ │ Metadata & Local WebSockets ▼ ┌────────────────────────────────────────────────────────────────────────┐ │ YOUR CUSTOM SAAS PORTAL / MOBILE APP (React/TS) │ │ - Executes text query: "Find man wearing a Dodger hat" │ │ - Maps cosine similarity rankings instantly against SQLite-VSS │ └────────────────────────────────────────────────────────────────────────┘ 2. Definitive Bill of Materials (BOM) & Equipment Prices Every component selected for this turnkey NVR appliance balances cost, performance, and long-term 24/7 durability. Below is the wholesale pricing structure to factor into product margin calculations.Turnkey AI NVR Appliance: Component Bill of Materials (BOM) Component Selected Hardware Model Mission-Critical Engineering Purpose Processor (CPU) Intel Core i5-12400 (6C/12T) Eliminates CPU rendering limits. Houses Intel UHD 730 QuickSync to handle high-density RTSP/H.264/H.265 ingestion at near 0% host CPU load. Graphics & AI GPU Nvidia GeForce RTX 5060 Ti (16GB) Blackwell Architecture Upgrade. Supports native FP4 precision processing to execute description searches 2x faster, while the 16GB VRAM cushion prevents Out-Of-Memory (OOM) crashes during concurrent camera alerts. System Memory 32GB (2 x 16GB) DDR5-5200 RAM Critical cushion to support multi-stream pipelines and host the live uncompressed RAM-disk buffer. Database Storage 500GB PCIe 4.0 NVMe SSD High-read endurance profile. Isolated exclusively to the Linux OS core, the Frigate framework databases, and fast local vector indices. Surveillance Storage 8TB Western Digital Purple Pro HDD Built specifically for surveillance workloads to withstand continuous 24/7 write loops over multiple years without failing. Power Supply Unit Seasonic Focus GX-650 (80+ Gold / ATX 3.1) Upgraded Component. Replaces failure-prone budget power supplies with heavy-duty Japanese solid capacitors rated for continuous 24/7 high-uptime workloads. Chassis & Case Micro-ATX Short-Depth Chassis Compact design built for server closets or office desk environments, utilizing positive pressure filtering to isolate internal parts from dust. Cooling Array Noctua IndustrialPPC 120mm PWM Fans Replaces fragile consumer case fans with structural, dust-and-waterproof industrial cooling units to prevent physical lockups. 3. Developer Configuration & Optimization Scripts 🛡️ Core Hardware Protection: 100% RAM disk for Live Motion Cache To prevent continuous video loops from destroying the system's primary NVMe SSD within months, configure Frigate to write all live temporary cache frames directly into system RAM via tmpfs. The system will only write to the physical hard drive once a permanent clip event is confirmed. Add this exact layout configuration into your deployment docker-compose.yml: yaml version: "3.9" services: frigate: container_name: frigate privileged: true # Required for hardware acceleration access restart: unless-stopped image: ghcr.io/blakeblackshear/frigate:stable shm_size: "256mb" # Allocated memory per camera connection footprint volumes: - /etc/localtime:/etc/localtime:ro - /opt/nvr/config:/config - /mnt/storage/surveillance:/media/frigate # Points directly to WD Purple HDD # CRITICAL: Mounts a 1GB temporary RAM-disk path to prevent NVMe wear - type: tmpfs target: /tmp/cache tmpfs: size: 1000000000 # 1GB maximum ceiling allocation ports: - "5000:5000" - "8971:8971" # Authenticated UI Access Port runtime: nvidia # Passes Blackwell framework to container layer Use code with caution. ⚡ GPU Under-Volting and Fan Curve Management (Linux Host Daemon) Consumer graphics cards will wear down their physical cooling bearings if left unmanaged. Implement this native system bash automation script on the underlying Ubuntu base image to lock power draws, lower temperatures, and set a low, steady fan speed. Create a system service deployment file at /etc/systemd/system/nvidia-nvr-tuner.service: bash #!/bin/bash # Enable persistence mode across the GPU matrix nvidia-smi -pm 1 # Under-volt/Cap maximum operational wattage limit to lower temperatures # Caps the Blackwell card to a steady 115W instead of spiking to 145W+ nvidia-smi -pl 115 # Enable manual fan manipulation control layers nvidia-settings -a "[gpu:0]/GPUFanControlState=1" # Force fans to run at a static, low speed (e.g., 40%) # Prevents aggressive RPM revving cycles that destroy fan bearings over years nvidia-settings -a "[fan:0]/GPUTargetFanSpeed=40" nvidia-settings -a "[fan:1]/GPUTargetFanSpeed=40" Use code with caution. 4. Final Developer Alignment Sheet Review this technical alignment sheet during your initial deployment sprint to verify that your system architecture is fully prepared to run semantic text searches locally: Technical Attribute Deployment Protocol Requirement Target Host Operating System Clean, Headless Ubuntu Server 24.04 LTS (No desktop GUI overhead). Camera Ingestion Compliance Strict ONVIF Core Protocol Compliance using WS-Discovery for automatic camera provisioning via the custom React portal. AI Processing Framework Nvidia TensorRT Execution Engine running natively on Blackwell tensor nodes. Vector Search Pipeline Method Ingested snapshot keyframes must pass through a local Jina-CLIP embedding variant via Docker container loops. Vector Index Data Layer Vector arrays must save into a local, high-speed SQLite-VSS database file living on the host NVMe SSD layer. SaaS Pipeline Bridge The custom cloud application webapp communicates with the edge box via Encrypted Local WebSockets or MQTT event endpoints. Long-Term System Stability A system crontab must fire an automated container teardown, RAM flush, and reboot script every Sunday at 3:00 AM. 5. Implementation Roadmap Action Items 🟩 Step 1: Base Image Architecture Spin up headless Ubuntu Server on the target Intel prototype hardware. Verify that Intel QuickSync hardware acceleration drivers are active (/dev/dri/renderD128). Install nvidia-container-toolkit to pass the local Blackwell processing layer through to Docker. 🟨 Step 2: Ingestion & Vector Testing Deploy Frigate using the tmpfs configuration file detailed above. Hook up an ONVIF-compliant IP camera stream and monitor system resources to verify that host CPU usage stays near 0%. Generate a collection of local image vectors using a test CLIP framework and measure search query response latency. 🟦 Step 3: Custom UI & SaaS Binding Construct a reverse proxy interface to safely map custom React app queries down to the local SQLite database. Validate that text queries (like "man wearing a Dodger hat") successfully fetch the correct video clips within milliseconds.