Quick Summary (TL;DR): For four modestly active IP cameras, start with a modern Intel mini PCAliExpress price such as the Beelink EQ13 N100 recommended in Frigate’s own hardware guide, or a compatible N100/N150 system with 8–16GB RAM, working Intel video decoding and an appropriately sized recording drive. For eight cameras, 16GB RAM and a documented Intel GPU/OpenVINO path are a more comfortable starting point; a Core Ultra 5 mini PCAliExpress price is worth considering if several cameras face busy areas or you want extra AI features. For 16 cameras, plan a 32GB-capable, expandable Intel system, local recording storage, a reliable wired network and measured detector headroom; add a supported accelerator or stronger GPU if activity demands it. Camera count alone is not a performance guarantee. Frigate processes a low-resolution detection stream and can record a separate full-quality stream, so the actual workload depends on simultaneous motion, resolution, codec, detection model, hardware decoding and retention. Do not automatically buy a Coral for a new build: current Frigate documentation no longer recommends it as the default new-installation choice. This is a documentation-led buying guide, not our own camera-count benchmark.
Best Frigate Hardware for 4, 8 or 16 Cameras: The Short Answer
| Installation | Practical starting hardware | Memory target | Detection route | What could force an upgrade |
|---|---|---|---|---|
| 4 cameras | Beelink EQ13 N100 or comparable Intel AVX2-capable mini PCAliExpress price | 8–16GB | Intel iGPU with OpenVINO; check supported model | Several busy views, high-res detect streams, additional AI enrichments |
| 8 cameras | N150/N100 for low activity; Core Ultra 5-class mini PCAliExpress price for heavier use | 16GB | OpenVINO GPU; NPU on supported Core Ultra systems | Overlapping motion, face recognition/LPR, multi-user live views |
| 16 cameras | Core Ultra 5-class 32GB-capable host or expandable business workstation | 32GB starting budget, adjusted to workload | Intel GPU/NPU or compatible Hailo; verify inference queue and decode capacity | Busy public-facing views, demanding models, 4K detection, heavy enrichments |
These are procurement tiers and test starting points, not vendor-certified “supports 16 cameras” figures. Frigate’s official planning guide separates low activity (roughly 1–6 cameras), moderate activity (6–12) and high activity (12 or more), and explicitly says simultaneous motion matters. Its current recommended-hardware page favours the Beelink EQ13 and warns of known compatibility issues with the EQ14; do not substitute EQ14 merely because it has a newer N150 processor.
We are comparing the Frigate host and its supporting recording/acceleration hardware, not ranking camera brands or giving a full Frigate installation tutorial. For a combined Home Assistant and media server, our mini PC plus NAS architecture guide explains when separating compute and storage is preferable.
Why “How Many Cameras?” Is the Wrong First Specification
An NVR does several jobs at once: receive network streams, decode detection frames, look for motion, run an object-recognition model on selected regions, manage live viewing, and retain video on disk. Those operations stress different hardware. A fast neural accelerator will not fix a system that is spending most of its CPU time decoding unsuitable 4K streams; a fast SSD will not fix a slow object detector.
| Workload | Main limiting resource | What to verify before purchase |
|---|---|---|
| Receiving RTSP streams | Network stability, camera firmware, connection count | Wired PoE, stream reliability, VLAN/firewall plan |
| Decoding detection substreams | GPU media engine and driver/device mapping | Supported H.264/H.265 profile, hardware-acceleration preset |
| Motion analysis and tracking | CPU, frame size and detect frame rate | Correct detect resolution, realistic 5fps target |
| Object inference | GPU/NPU/TPU, model size, concurrent motion | Detector latency, queueing and actual busy-scene behaviour |
| Continuous recording | Drive capacity, sustained write reliability | Measured bitrate × cameras × retention, spare space |
| Live view and enrichment | GPU/CPU/RAM, browser/client compatibility | Concurrent viewers, LPR, face recognition, semantic search |
Frigate’s video-pipeline explanation is particularly useful: it decodes the detection stream, finds motion regions, and sends selected crops to the model. Five detection frames per second from 16 cameras do not mean 80 neural-network inferences every second. Not every frame contains an object requiring a fresh inference, and one busy frame can create several regions. A “100 inferences per second” detector claim cannot be translated directly into a guaranteed camera count.
The Best Entry-Level Choice: Beelink EQ13 N100 for Four Cameras
The Frigate project currently calls the Beelink EQ13 a favourite compact server because its Intel N100 is efficient and the dual network ports can support a segregated camera network. The recommendation describes several 1080p cameras with low-to-medium activity, not an unconditional four-camera ceiling. This makes it a sensible first machine for a driveway, front door, back garden and side gate when each camera offers a useful detection substream.
Before ordering, verify the exact N100/N200 configuration, RAM, SSD, both Ethernet ports, regional mains lead, Linux compatibility, return rights and warranty. The EQ13 has appeared in different regional bundles and is not necessarily readily available new in October 2026. Buying an overpriced, discontinued configuration purely because it appears in a recommendation is poor value; a documented equivalent Intel mini PCAliExpress price can be sensible after testing the same Frigate requirements.
- CPU: Intel N100 is a four-core, four-thread low-power part; it supports AVX2, which matters because Frigate’s current planning guide requires AVX and AVX2.
- GPU: make the integrated graphics available to Frigate for video decoding and OpenVINO inference; do not assume the container sees it automatically.
- RAM: 8GB can work for a straightforward four-camera build; choose 16GB if you want sensible margin for Docker, the OS and occasional extra services.
- Storage: a small boot SSD is not a two-week archive; budget a separate, correctly sized recording disk.
- Network: two NICs are useful for isolation but do not replace VLAN policy, strong camera credentials and an appropriately sized PoE switch.
Important EQ14 caveat: Frigate’s official recommended-hardware page specifically says to avoid the Beelink EQ14 for now because of known compatibility issues. That warning takes priority over the attractive N150 specification or any earlier general-purpose mini-PC recommendation. Check the Frigate project’s latest guidance before treating a firmware update as a cure.
Eight Cameras: When to Choose N150 and When to Step Up to Core Ultra 5
Eight cameras is the point where scene activity starts to matter more than an appealing low-power processor name. Eight quiet indoor or residential cameras with sensible 5fps substreams may run well on a modern Intel N100/N150 host. Eight cameras facing a busy pavement, entrance, car park or overlapping driveways can create sustained motion and inference requests. Recording bitrate also doubles compared with four equivalent cameras, even if inference remains light.
An Intel N150 system can be a good fit if the exact mini PC has stable Linux graphics support, 16GB of RAM, enough recording storage and verified cooling. Frigate’s published OpenVINO table lists example inference timings for N100 and N150, but these are model- and environment-specific examples, not a direct camera-capacity certification. A particular N150 chassis may perform worse than another if drivers, BIOS, thermals or power limits differ.
For a heavier eight-camera system, the ASUS NUC 14 Pro with Core Ultra 5 125H is an identifiable higher-headroom alternative. ASUS’s UK specifications list Core Ultra 5 125H configurations, two DDR5 SO-DIMM slots, Intel graphics, 2.5GbE and M.2 SSD support. The exact kit may arrive without RAM or SSD. Intel’s NPU is potentially useful for Frigate object detection, but the host kernel/firmware and Frigate detector configuration must support it. Do not assume every “AI PC” feature is exposed to every container.
ASUS also notes that Intel Arc graphics branding on certain NUC 14 Pro variants requires two SO-DIMMs. If buying barebones for GPU-related work, use the exact UK SKU specification and plan the memory configuration accordingly. The NUC is a platform, not a ready-made NVR: recording drives, cooling placement, UPS and software setup remain your responsibility. See ASUS NUC 14 Pro official specifications.
Sixteen Cameras: Buy for Sustained Load and Recovery, Not a Headline CPU Score
At 16 cameras, a single mini PC can still be viable, but the design needs more discipline. The Frigate planning guide treats 12 or more cameras as a high-activity tier when multiple feeds may be busy. For a new build, favour a 32GB-capable Intel Core Ultra system or an expandable business desktop/workstation with accessible PCIe slots, documented GPU support, proper cooling and space for dedicated recording media. This is a procurement recommendation rather than a claim that 32GB is an absolute requirement.
The strongest reason to choose an expandable workstation over a sealed mini PC is not simply more CPU threads. It is the ability to add an Intel Arc-class GPU or supported accelerator, more local storage, additional NICs and serviceable cooling if the real workload grows. A compact NUC can be excellent when the integrated GPU/NPU meets your measured demand and the archive lives on reliable external storage. A tower may be better when you need several local surveillance HDDs and easy replacement.
Do not buy a discrete GPU merely because you have 16 cameras. First use the correct low-resolution detection streams, enable supported decoding, tune motion masks and measure the inference queue. Conversely, do not assume a single Coral or integrated GPU can sustain 16 busy 4K streams with face recognition and licence-plate recognition just because a forum post says “16 cameras works”. Different models, resolutions and event rates make that claim meaningless without a test plan.
| Buying route | Best reason to choose it | Expansion limit / risk | Our verdict |
|---|---|---|---|
| N100/EQ13 compact host | Low power, official Frigate preference for modest residential load | Limited RAM/drive space; needs real validation at 8+ cameras | Best first four-camera host |
| N150 mini PC (not EQ14) | Efficient Intel GPU/OpenVINO path; 16GB builds are common | SKU and driver compatibility; memory/SSD ceilings vary | Good quiet eight-camera candidate |
| ASUS NUC 14 Pro Core Ultra 5 | More compute/RAM, iGPU and possible NPU route | Barebones cost; limited internal 3.5-inch recording bays | Best compact headroom choice |
| Refurbished AVX2 business desktop | PCIe expansion, replaceable cooling, internal disks | Age, idle power, warranty, drive cages and GPU codec generation | Best expandable 16-camera value candidate |
| Raspberry Pi 5 + supported Hailo AI HAT+ | Compact, efficient accelerator experimentation | Video decode/storage/network architecture needs careful proof | Specialist small deployment, not default 16-camera NVR |
Intel OpenVINO, Hailo and Coral: Which Detector Belongs in the Budget?
A detector accelerates object inference; it does not automatically accelerate RTSP video decoding, disk writes, or every AI feature. Frigate currently supports Intel OpenVINO on suitable integrated GPUs, Arc GPUs and NPUs; Hailo-8/Hailo-8L; Google Coral Edge TPU; and several other backends. Software support and the chosen neural model are as important as advertised TOPS.
| Detector option | Why it makes sense | Compatibility check | Buying warning |
|---|---|---|---|
| Intel OpenVINO GPU | Reuse an Intel iGPU already in the host | Host drivers, render device mapping, model compatibility | Integrated graphics may be shared with video decoding |
| Intel OpenVINO NPU | Free GPU resources for decoding and other workloads on supported Core Ultra hardware | Host kernel, NPU firmware and Frigate version | NPU label on a product page is not proof it works in your VM |
| Hailo-8 / 8L | Dedicated inference on compatible M.2/PCIe hosts | Physical keying, PCIe lanes, model/runtime, cooling, host mapping | Do not assume an ordinary M.2 SSD socket fits an AI module |
| Google Coral USB/M.2 | Useful for an existing setup or constrained low-power host | Supported Edge TPU model, USB/PCIe passthrough, driver | Frigate no longer recommends Coral as the default new purchase |
| Discrete Intel Arc GPU | More headroom for busy cameras and advanced models | Slot, PSU, Linux kernel/driver, video-decoding support | Extra heat, power and enclosure cost |
The Frigate hardware page gives illustrative MobileNetV2 inference times of about 15ms for N100 and N150, and model-specific figures for Intel NPUs and Arc GPUs. At 15ms, 1,000 ÷ 15 ≈ 67 is a theoretical, serial inference-per-second ceiling before scheduling and other overhead. It is not 67 guaranteed object detections or a promise about 16 cameras. A more demanding YOLO model may be slower. Monitor queueing and actual latency under your chosen model, not a different model’s marketing number.
For Hailo, Frigate currently supports Hailo-8 and Hailo-8L and describes default YOLOv6n models; check the current detector documentation for installation and configuration. Raspberry Pi’s original AI Kit is no longer in production: its AI HAT+ documentation identifies the Hailo-8L and Hailo-8 variants. Do not confuse the newer AI HAT+ 2 hardware with a Frigate-verified drop-in replacement for Hailo-8/8L.
Detection Stream Settings Can Save More Than a Faster CPU
Frigate recommends separate camera streams where possible: a lower-resolution stream for detection and a higher-resolution stream for recordings. Its current camera-setup guide says the default 5 detection frames per second is right for almost all ordinary scenes, while a recording stream might be configured at 15fps in the camera firmware. Set the camera’s actual substream frame rate where possible; decoding a high-fps stream only to discard frames wastes resources.
| Setting | Starting point | Reason | When to change it |
|---|---|---|---|
| Detect stream | Camera-native substream, often 640×360 or 640×480 | Lower decoded pixel count; adequate for normal approach scenes | Tiny/distant subjects need more detail; keep matching aspect ratios where possible |
| Detect fps | 5fps | Frigate default and normal recommendation | Only increase after debug-view evidence that fast cross-traffic is missed |
| Record stream | Full-quality main stream, e.g. 1080p/4MP/4K | Retains identification detail without decoding all pixels for inference | Choose bitrate/codec according to evidence and storage budget |
| Codec | H.264 for maximum compatibility | Broad browser and HA support | H.265 may reduce bitrate but limits browser/decoder compatibility |
| Motion masks | Mask clocks, moving trees and irrelevant regions carefully | Reduces false work and alert noise | Never mask the real object approach path |
| Camera connectivity | Wired Ethernet/PoE where possible | Stable streams and predictable power | Wi-Fi only where wired installation is genuinely impractical |
For example, four 4K cameras do not require decoding four 4K images at 30fps for detection if the cameras supply suitable substreams. The full-quality main feeds can be recorded without re-encoding. Conversely, a cheap camera that offers only one high-resolution stream can create a much more demanding host workload than a more expensive dual-stream camera.
Use the current Add Camera Wizard to probe streams and assign detect/record roles. The wizard tests codecs, resolutions and URLs. A fixed recipe copied from an old blog may not match your camera firmware or current Frigate UI.
Worked Sizing Example: Camera Count, Activity and Inference Demand
Here is an explicit planning model, not a benchmark. Assume all cameras provide a 640×360 detection substream at 5fps, a separate 4Mbps average recording main stream, H.264 video, and a modest default object model. The motion activity pattern below is a design assumption: in the four-camera home, one or two feeds are busy at once; in the eight-camera home, three or four; in the sixteen-camera site, six to eight during a busy period.
| Scenario | Total detection frames decoded/s at 5fps | Assumed simultaneous busy feeds | Illustrative inference requests/s if each busy feed triggers 3 regions/s | What the example demonstrates |
|---|---|---|---|---|
| 4 cameras | 20 | 1–2 | 3–6 | Decoding is continuous; inference can remain intermittent |
| 8 cameras | 40 | 3–4 | 9–12 | Double the feeds need not double inference unless activity rises |
| 16 cameras | 80 | 6–8 | 18–24 | Busy scenes may be feasible with fast detection, but must be measured |
The “three regions per busy camera per second” figure is invented solely as a transparent workload assumption, not measured Frigate behaviour. Real queues may spike much higher, especially with overlapping objects or poor motion masking. The calculation deliberately separates decoded frames from model inference requests. If a detector averages 15ms per inference, 24 serial requests/s would occupy roughly 360ms of every second for inference alone in this simplified model; decoding, image preparation, scheduling and burstiness still matter. Test a worst-case evening, rain or busy-road scene before approving the hardware.
A practical commissioning test is to run all cameras for several days, observe the busiest hour, check CPU and GPU utilisation, inference latency, dropped detection frames, restarts, storage writes and playback. If the model queue grows persistently or the host thermally throttles, first correct stream settings and then upgrade the bottleneck. A single five-minute quiet test proves very little.
Recording Storage: The 4Mbps Calculation for 4, 8 and 16 Cameras
For a constant average stream bitrate, recording storage is easy to estimate: GB per day ≈ Mbps × 10.8 for each camera, using decimal gigabytes. At an assumed 4Mbps average, one camera generates about 43.2GB/day. Actual bitrate varies with scene detail, codec, variable-bit-rate settings, night noise, audio and the camera’s rate-control behaviour. This is not a claim that every 1080p or 4K camera records at 4Mbps.
| Cameras × 4Mbps assumed | Network recording traffic | 7 days continuous | 14 days continuous | 14-day target with 20% capacity margin |
|---|---|---|---|---|
| 4 | 16Mbps | 1.21TB | 2.42TB | 2.90TB |
| 8 | 32Mbps | 2.42TB | 4.84TB | 5.81TB |
| 16 | 64Mbps | 4.84TB | 9.68TB | 11.61TB |
The 20% margin is a planning allowance, not a filesystem or Frigate rule. Reserve additional space for exports, thumbnails, the database, snapshots and filesystem overhead. For a 16-camera 14-day continuous system, a nominal 12TB recording volume might look adequate under this 4Mbps model but leaves little extra headroom once actual disk usable capacity, filesystem reserves and variable bitrate are considered. Check the real usable space and whether you need redundancy or independent backups.
For motion/event-only retention, do not apply an arbitrary “70% saving”. Measure actual retained segments after a representative week; Frigate records from the original stream without re-encoding and its recording policy determines which segments survive. Event-only storage can be far lower in quiet scenes, but it can approach continuous use in a busy street. Continuous recordings, event clips, snapshots and exported evidence have different retention behaviour.
A surveillance-rated HDD such as a current WD Purple or Seagate SkyHawk is a reasonable long-retention option, subject to the exact drive’s datasheet and warranty. SSDs are attractive for quiet installations and responsiveness, but select capacity and write endurance for the actual recording workload. A NAS is supported as a target, yet Frigate’s planning guide prefers local recording storage for reliability and latency; remote SMB/NFS storage adds another failure dependency.
Network, PoE and Security: Often the Missing Part of a “Mini PC” Quote
Sixteen 4Mbps recording streams total only 64Mbps in the simple example, well below a healthy 1GbE link. That does not mean any cheap switch is automatically suitable. You still need PoE power budget, enough physical ports, a stable uplink, correct cabling, multicast/broadcast behaviour where relevant, and headroom for live viewers, camera management and NAS traffic. A 2.5GbE uplink may help a busy mixed-use network, but 2.5GbE is not mandatory merely because you have 16 cameras.
For a camera network, favour a managed PoE switch where VLAN isolation is needed. Deny direct internet access to cameras unless an explicit, necessary feature requires it; allow the Frigate host to reach the camera streams and time service as appropriate. Do not expose unauthenticated RTSP or Frigate administrative interfaces directly to the public internet. Use unique camera passwords and maintain firmware. If your network also serves a NAS, our 2.5GbE switch comparison explains when faster links and VLAN management are worth paying for.
Budget the PoE switch and its power supply separately from the server. Eight cameras at 8W each would imply 64W of hypothetical camera power before switch overhead and reserve; that is a sizing example, not a specification for a particular camera. Use each camera’s manufacturer maximum draw (including IR LEDs and heaters) and the switch’s total PoE budget, not just its port count. A 16-port switch with a 120W shared budget can be inadequate for 16 high-draw cameras.
Mini PC vs Raspberry Pi 5 vs NAS vs Refurbished Desktop
| Host type | Strongest use case | Where it disappoints | Buying advice |
|---|---|---|---|
| Intel mini PC | Quiet four/eight-camera Frigate and Home Assistant host | Few internal 3.5-inch bays; limited PCIe | Best default when recording can live on a separate supported disk |
| Raspberry Pi 5 + Hailo AI HAT+ | Low-power, technically interesting small build | Less forgiving for many high-res streams and archive expansion | Choose after verifying actual Frigate video decoding and recording path |
| NAS running Frigate container | Existing NAS has spare CPU/RAM and supported acceleration | Container device mapping, shared disk load and downtime coupling | Verify exact NAS CPU/OS/Frigate support; not all NAS appliances are equal |
| Used business desktop | Expandable 16-camera host with PCIe and local HDDs | Variable age, noise, power and warranty | Best when you need drive bays, serviceability and a real upgrade path |
| Core Ultra NUC/compact workstation | Mixed NVR + other services with GPU/NPU headroom | Higher total system cost; storage remains external | Buy for workload growth, not solely camera count |
A Raspberry Pi 5 with Hailo AI HAT+ can accelerate inference, but it does not magically gain a desktop-class video decoder, multiple drive bays or an enterprise recording subsystem. Likewise, running Frigate inside a NAS or Proxmox VM may be convenient, but the GPU/accelerator must be passed through correctly. If you also host Home Assistant, see our Home Assistant OS vs Docker vs Proxmox guide for the management and isolation trade-offs.
A small ESP32-CAM can be useful for an inexpensive snapshot or maker experiment, but do not treat a Wi-Fi ESP32-CAM board as equivalent to a wired, dual-stream surveillance IP camera for a reliable 16-camera NVR. Stream stability, night performance, weatherproofing and camera firmware matter more than the microcontroller’s price.
Complete Ownership Cost: What the Hardware Listing Leaves Out
For a buying decision, compare the complete five-year cost of a working and recoverable system, not the bare mini-PC sticker price. We do not publish unverified October 2026 street prices, so the purchase columns below are intentionally quote-driven. Obtain live UK/EU quotes for the exact regional SKU, including VAT, warranty, returns and the included RAM/SSD. The cost of cameras themselves is excluded if you already own them; include it if you are building from scratch.
| Cost component | Four-camera build | Eight-camera build | Sixteen-camera build |
|---|---|---|---|
| Compute host | EQ13 N100 or verified equivalent, quoted complete | N150/NUC Core Ultra 5, quoted complete | Core Ultra 5/workstation, quoted complete |
| RAM + boot SSD | Confirm whether 8–16GB and SSD included | Confirm 16GB and SSD, add missing parts | Price 32GB, boot SSD and expansion path |
| AI accelerator | Often no extra hardware if Intel GPU works | Only if measured OpenVINO headroom is insufficient | Optional Hailo/Arc/NPU path after testing |
| Recording storage | Capacity from retention calculation | Capacity from retention calculation | Multiple disks/enclosure or local array if needed |
| PoE and cabling | PoE ports, injectors, cables, switch power | 8-port/greater PoE budget, cabling | 16+ ports, sufficient PoE watts, uplink |
| Power protection | UPS for host, switch, router/ONT | UPS sized for combined measured load | UPS + orderly shutdown, potential battery replacements |
| Data protection | Off-host configuration backup, export policy | Off-host config, incident export procedure | Independent config/critical-footage backup and restore tests |
| Five-year running cost | Measure wall draw; formula below | Measure wall draw; formula below | Measure entire host + drives + PoE stack |
A useful worksheet is five-year total = host + RAM + boot SSD + recording disks/enclosure + optional detector/GPU + PoE switch/injectors + cabling + UPS + replacement batteries + backup media + delivery/taxes + electricity. Add any subscription only when you have verified you actually need it. Frigate itself is open source; do not invent a referral payment for Frigate or Frigate+ features.
For electricity, the Ofgem 1 October–31 December 2026 Great Britain Direct Debit average is 26.32p/kWh. This is a dated illustration, not a five-year tariff forecast or the correct rate for every UK household. The standing charge is not added to the incremental NVR cost if you already pay it for the property.
| Illustrative measured whole-system average | kWh/year at 24/7 | One year at 26.32p/kWh | Five years at unchanged rate |
|---|---|---|---|
| 15W | 131.4 | £34.58 | £172.92 |
| 25W | 219.0 | £57.64 | £288.20 |
| 40W | 350.4 | £92.23 | £461.13 |
| 60W | 525.6 | £138.34 | £691.69 |
The watt figures are illustrative metered-power inputs, not tested draws of any named PC. A real 16-camera stack may consume considerably more when disks, PoE cameras, switches and UPS losses are included. For a fair comparison, meter the same boundary in each design. An N150 processor’s nominal base power is not the same as the mini PC’s wall consumption; similarly, a UPS rated 900VA does not constantly consume 900W. At 26.32p/kWh, each additional continuous 10W costs approximately £23.06 per year, or £115.28 over five years if that unit rate stayed unchanged.
If you need a UPS, check both VA and watts, the correct outlet type and tested shutdown support. Our UPS for Home Assistant, router and NAS guide covers those purchase details. Cameras powered from a PoE switch must be counted in the power budget and backup runtime if you expect recording to continue during an outage.
A Sensible Configuration and Commissioning Checklist
Frigate’s current camera wizard is the preferred way to create camera entries. The following is a minimal illustrative YAML fragment showing the roles of two streams on a single camera; it is not a complete production configuration, and the RTSP paths are placeholders. Check the exact current Frigate version and the camera vendor’s RTSP URLs before deployment.
mqtt:
enabled: false
detectors:
ov:
type: openvino
device: GPU
ffmpeg:
hwaccel_args: preset-vaapi
cameras:
front_door:
ffmpeg:
inputs:
- path: rtsp://CAMERA_USER:CAMERA_PASSWORD@192.0.2.10/substream
roles: [detect]
- path: rtsp://CAMERA_USER:CAMERA_PASSWORD@192.0.2.10/mainstream
roles: [record]
detect:
enabled: true
fps: 5
record:
enabled: true
continuous:
days: 7
This assumes an Intel GPU visible to the Frigate container and a camera with separate main and substreams. The 192.0.2.10 address is a documentation-only example, not a real camera. Credentials in a real installation should be protected rather than copied into public documentation. The example retains continuous footage for seven days; adjust retention only after measuring storage. The hardware-acceleration preset must match your driver and codec. Frigate’s video decoding guide and FFmpeg preset reference explain the supported choices.
- Inventory every camera: write down main/substream resolution, H.264/H.265 codec, bitrate, frame rate, audio and exact RTSP/ONVIF support.
- Set a retention target: continuous or event-based, days to keep, exports and privacy/legal obligations; calculate usable recording capacity.
- Choose the host: verify AVX2, Linux support, cooling, GPU/NPU access, memory expansion, storage interface and vendor warranty.
- Configure one camera: use the Frigate wizard, confirm decode acceleration and detector operation, then expand to all cameras.
- Stress-test: exercise several simultaneous motion zones, enable representative live views and any enrichments, and monitor dropped frames, queue latency and CPU/GPU usage.
- Verify storage and resilience: check 24-hour write totals, retention deletions, disk health, UPS shutdown and recovery after an unexpected power loss.
- Protect evidence: confirm exported clips can be played independently and critical configuration/backups exist off the NVR disk.
Common Buying Mistakes That Waste Money
- Buying only by camera count: four 4K single-stream cameras can be harder to decode than eight properly configured dual-stream cameras.
- Using CPU object detection by default: Frigate advises hardware-accelerated inference; the generic CPU detector is primarily a fallback/test option.
- Buying a Coral because an old guide says it is mandatory: Frigate’s current guidance has moved on; check OpenVINO, Hailo and other supported paths.
- Ignoring the EQ14 warning: the Frigate project currently flags compatibility issues with that exact Beelink model.
- Assuming GPU passthrough: Proxmox, Docker, Home Assistant apps and NAS containers do not expose identical devices or permissions.
- Buying the smallest SSD: even 4Mbps × 8 cameras produces roughly 346GB of continuous recordings per day.
- Forgetting camera PoE load: a powerful mini PC cannot help if the PoE switch browns out when IR LEDs activate.
- Relying on a quiet afternoon test: night noise, rain, headlamps, moving foliage and busy streets may substantially increase event volume.
- Skipping restore testing: RAID, a UPS and a local snapshot do not replace a recoverable off-host configuration backup.
Frequently Asked Questions
Can an Intel N100 really run Frigate with eight cameras?
It may, especially with low-activity 1080p cameras, appropriate 5fps detection substreams and working Intel video acceleration. Frigate’s official recommendation for the EQ13 is deliberately phrased as “several 1080p cameras with low-medium activity”, not a guaranteed eight-camera result. Validate all eight real feeds and peak motion before committing to the host.
Do I need a Coral TPU for Frigate in 2026?
No. Frigate supports Intel OpenVINO, Hailo and other accelerators, and its current hardware page explicitly says Coral is no longer the preferred default for new installations. Existing Coral devices remain useful in supported setups, particularly where power use is constrained.
Does a 16-camera Frigate system need 10GbE?
Not from camera count alone. At the article’s assumed 4Mbps per camera, 16 streams total 64Mbps before protocol overhead, which is well below Gigabit Ethernet capacity. Faster networking may matter for multiple live viewers, NAS writes, backup transfers or other shared traffic, but stable wired PoE and switch power budget come first.
How much RAM should I buy for Frigate?
Frigate’s planning documentation describes 4GB as a basic minimum, 8GB as a minimum when using enrichments, and 16GB as recommended for most larger setups, particularly eight or more cameras. We would start with 8–16GB for four, 16GB for eight and a 32GB-capable host for a demanding 16-camera deployment, then size against actual workloads. Additional services and VMs need their own memory allocation.
Is a Raspberry Pi 5 with Hailo better than an N100 mini PC?
Not automatically. Hailo can make inference efficient, but the rest of Frigate still needs to decode video, handle networking and write recordings. For a small specialist setup, Pi 5 plus supported Hailo hardware can be attractive; for an ordinary multi-camera NVR, the project’s current Intel mini-PC recommendations and x86 video-acceleration path are often simpler. Compare complete hardware and storage costs, not just accelerator TOPS.
Should I record to an SSD or a surveillance HDD?
Use an SSD where silence, latency and compactness matter and the drive’s endurance and capacity are suitable. Use surveillance-rated HDDs for large, cost-effective archives, especially 8–16-camera continuous recording. Frigate’s planning guide discusses both. Separate critical exports and configuration backups from the main recording volume.
Will an AI accelerator improve 4K live viewing?
An object detector is not a video-decoding engine. It can speed up recognition of objects in motion regions, while an Intel/AMD/NVIDIA media engine handles decoding and possibly scaling or encoding. If 4K live view is stuttering because of network or codec incompatibility, buying a Coral or Hailo module may change nothing.
Our Verdict: What We Would Actually Buy
For four residential cameras, begin with a properly supported Intel N100 mini PC—ideally the Frigate-recommended EQ13 if available at a sensible verified price—or an equivalent AVX2-capable Intel host, 8–16GB RAM, an appropriate recording drive and reliable PoE. For eight cameras, an N150-class host with 16GB is plausible when activity is modest, but a Core Ultra 5 platform offers more useful margin for busy scenes and future features. For sixteen cameras, prioritise an expandable 32GB-capable host, storage and power resilience, and a documented GPU/NPU/accelerator path that survives a real peak-load test. We would spend on stable camera streams, storage and backups before buying an oversized AI accelerator.
The crucial purchasing question is not “Which processor claims 16 cameras?” It is “Can this complete, serviceable system decode my streams, keep up with the busiest real scene, retain footage for the required period and recover safely after failure?” That is the standard a useful Frigate NVR should meet.
Related ESP32 and Home-Server Guides
- Mini PC Plus NAS vs a NAS-Only Home Server — decide where to place Frigate compute and recordings.
- Home Assistant OS vs Docker vs Proxmox — understand virtualisation and device-passthrough overhead.
- Best 2.5GbE Switches for a Smart Home and Home Lab — network choices when cameras share a busy LAN.
- Best UPS for Home Assistant, Router and NAS — safe shutdown and power sizing.
- ESP32-CAM Pinout: Safe GPIOs, Boot and Camera Pins — for experimental camera and snapshot projects.
Datasheets and Official Technical Sources
- Frigate — Recommended hardware (includes EQ13 preference, EQ14 warning, supported detectors and example inference figures).
- Frigate — Planning a New Installation (activity tiers, AVX2, RAM and storage).
- Frigate — Camera setup (5fps detection default, streams and codec choices).
- Frigate — Object detectors (OpenVINO, Hailo, Coral and configuration).
- Frigate — Hardware video decoding (Intel and other GPU acceleration).
- Frigate — Recording and retention (storage behaviour and export retention).
- ASUS UK — NUC 14 Pro technical specifications (Core Ultra 5 variants, RAM, networking and storage).
- Raspberry Pi — AI HAT+ documentation (Hailo variants and hardware differences).
- Ofgem — Electricity unit rates, 1 October to 31 December 2026 (dated UK electricity cost example).
Editorial verification: 10 October 2026. Hardware guidance is based on official documentation and specification comparisons, not a hands-on test. Model availability, UK/EU prices, firmware support, detector compatibility and energy tariffs can change; recheck the exact SKU before buying.