Best Frigate Server Build 2026: NVR Hardware for AI Object Detection

Introduction

If you are building a security camera system that actually understands what it sees—distinguishing a person from a falling leaf at 3 AM—you need a Frigate server built for AI object detection. The right hardware choice, particularly between a Google Coral TPU, an Intel CPU with Quick Sync, or a dedicated NVIDIA GPU, determines whether your system runs reliably on 15 watts or requires a 300-watt power budget. This guide covers three build tiers for 2–4 cameras, 6–10 cameras, and 10+ camera setups, with exact hardware picks, storage sizing, and network planning so you can make a confident purchase today.

Frigate Hardware Requirements: Why CPU-Only Detection Fails Past 2 Cameras

Frigate offloads object detection from the main CPU to a dedicated accelerator. Running detection purely on a modern CPU like an Intel i5-12400 might handle one or two 1080p streams at 5–10 frames per second (FPS), but add a third stream and inference latency spikes past 100ms, causing missed detections and high CPU load (often 70–90%).

For anything beyond a single test camera, you need either a Google Coral TPU, an Intel processor with integrated graphics (Quick Sync), or a discrete NVIDIA GPU. The accelerator handles the neural network inference—the actual “is that a person or a shadow?” math—while the CPU manages stream decoding, recording, and the web interface.

Google Coral TPU vs GPU Detection: Core Trade-Offs

The Google Coral TPU (a USB or M.2 accelerator) is the most power-efficient and cost-effective option for Frigate. It delivers about 4 TOPS (trillions of operations per second) per chip, enough for 6–8 camera streams at 5 FPS detection. A dual-Coral setup can handle 12–16 streams. The catch? Coral availability fluctuates, and it only handles inference—it cannot decode video streams.

GPU-based detection (using Intel Quick Sync or an NVIDIA card) combines stream decoding and inference in one device. Intel Quick Sync on a 12th-gen or newer i3/i5 handles 4–6 streams with decent accuracy, while an NVIDIA GTX 1650 or Tesla P4 can manage 10+ streams but draws significantly more power.

RAM and Storage Scaling by Camera Count

Frigate uses RAM for event buffers and decoded frames. For 2–4 cameras, 8GB of RAM is sufficient; for 6–10 cameras, 16GB is recommended; for 10+ cameras, 32GB ensures smooth operation, especially if you run additional services like Home Assistant alongside Frigate.

Storage is the bigger variable. Each camera recording 24/7 at 1080p with H.265 encoding consumes roughly 50–100 GB per day. At 4K, that jumps to 150–300 GB per day. Retention periods scale linearly—30 days of footage for four 1080p cameras requires about 6–12 TB.

Network and Bandwidth Planning for Multiple RTSP Streams

Each camera streams an RTSP feed to the Frigate server. A single 1080p camera at 15 FPS uses about 4–8 Mbps. Four cameras = 16–32 Mbps, well within a gigabit LAN. However, if you also stream to a separate NVR or view live feeds via the Frigate web interface, bandwidth doubles. For 8 cameras at 4K, expect 40–80 Mbps total—still fine for gigabit Ethernet, but Wi-Fi can struggle.

Always use wired Ethernet for cameras. PoE switches simplify power and data in one cable. For the Frigate server itself, a dedicated gigabit port (or better, a 2.5GbE port if you have 10+ cameras) prevents bottlenecks.

Build Tier 1: 2–4 Cameras (Budget Setup)

This tier targets a small home or apartment with 2–4 1080p cameras. The goal is low power, low cost, and silent operation.

• CPU: Intel N100 (6W TDP, includes Quick Sync for decoding)
• RAM: 8GB DDR4 or DDR5
• Storage: 1x 4TB SSD or HDD (enough for 30–60 days of events)
• Accelerator: Google Coral USB (M.2 if motherboard has slot)
• Case/PSU: Mini PC (e.g., Beelink S12 Pro) or custom ITX build with 60W PSU
• OS: Debian 12 + Docker or Proxmox LXC
• Estimated Cost: $200–$350

This build idles around 8–12W. The Coral handles detection for up to 4 streams easily. The N100’s Quick Sync decodes the video streams, keeping CPU load under 20%.

Build Tier 2: 6–10 Cameras (Mid-Range Workhorse)

For a larger home or small office with 6–10 cameras (mix of 1080p and 4K), you need more CPU grunt for decoding and a dual-accelerator option for detection.

• CPU: Intel i5-12400 or i5-13500 (65W TDP, strong Quick Sync)
• RAM: 16GB DDR4
• Storage: 2x 8TB HDD in RAID 1 (mirror) for 8TB usable, or single 12TB for event-only recording
• Accelerator: 2x Google Coral M.2 (dual TPU for up to 16 streams) OR 1x Coral + Intel iGPU
• Case/PSU: Fractal Design Node 804 or Jonsbo N2 with 250W PSU
• OS: TrueNAS Scale (for ZFS storage) or Unraid (for flexibility) — see our TrueNAS vs Unraid guide for comparison
• Estimated Cost: $600–$900

Power draw sits around 30–50W idle, 60–80W under load. The dual Coral setup provides redundancy—if one fails, detection continues on the second.

Build Tier 3: 10+ Cameras (Heavy-Duty NVR)

For 10–20 cameras at 4K resolution with 24/7 recording and 30+ day retention, you need enterprise-grade hardware.

• CPU: Intel i7-13700 or AMD Ryzen 7 7700 (for high decode throughput)
• RAM: 32GB DDR5
• Storage: 4x 12TB or 16TB HDD in RAID 5 (ZFS RAIDZ1) for 36–48 TB usable
• Accelerator: NVIDIA Tesla P4 (75W, 5.5 TOPS) or Intel Arc A380 (for AV1 decode)
• Network: 2.5GbE or 10GbE NIC
• Case/PSU: Fractal Design Define 7 XL or Supermicro SC846 with 500W+ PSU
• OS: Proxmox (virtualize Frigate + other services) or Ubuntu Server + Docker
• Estimated Cost: $1,500–$2,500

This build idles at 60–80W and draws 150–250W under full load. The NVIDIA GPU handles both decoding and inference for 15+ streams with sub-30ms latency.

Software Setup: Frigate on Docker or LXC

Frigate runs best in Docker on a Linux host (Debian, Ubuntu, or Proxmox LXC). The official container includes all dependencies. Key steps:

Which Should You Choose: Coral, Intel iGPU, or NVIDIA?

For most homelab setups (2–10 cameras), a single Google Coral TPU paired with an Intel CPU (N100 or i5) is the best balance of cost, power, and performance. It delivers reliable sub-30ms detection on 6–8 streams while drawing under 15W for the accelerator itself.

If you exceed 10 cameras or need 4K detection at high frame rates, switch to an NVIDIA GPU (Tesla P4 or RTX 3050) for combined decode and inference. The power cost is higher, but the performance ceiling is much higher.

Intel Quick Sync alone is only viable for 2–4 cameras on a tight budget where a Coral is unavailable—it works but detection accuracy and latency suffer compared to a dedicated TPU.

Frequently Asked Questions