Best Hardware for Frigate NVR (2026): Requirements and Recommended Specs

By LK Wood IV · Published 2026-07-30 · Updated 2026-10-11 · ~12 min read · St. Louis County, MO

Minimum specifications such as 4GB of RAM, AVX2 support, and an accelerator do not size the whole machine. Camera count, resolution, frame rate, retention, and the available detector decide how much hardware the system needs.

The table below sizes the host by camera count, with every requirement traced to Frigate’s own documentation.

The three gates before you spend anything

AVX and AVX2, or nothing works. Frigate’s planning guide requires “a CPU with AVX + AVX2 instructions.” Most modern processors carry both. The exception is exactly the hardware people buy for a cheap NVR. Frigate spells it out: “Intel Celeron and Pentium models prior to the 2020 Tiger Lake generation typically lack AVX.” That cheap used mini-desktop with a Celeron is a dead end. Check the specific model on Intel’s ARK before you buy, not after.

Hardware decoding, which is not the same as detection. Frigate’s video decoding docs say “it is highly recommended to use an integrated or discrete GPU for hardware acceleration video decoding in Frigate,” because pulling apart camera streams is the CPU-heavy half of the workload. This is the part people under-buy: they budget for an AI accelerator and forget that decoding eats a CPU alive without an iGPU.

RAM, in three published tiers. Frigate gives real numbers rather than a vague “more is better”:

Note that Frigate also holds recording cache in /dev/shm, sized separately from the above. Our Frigate NVR storage calculator computes that value alongside the disk estimate, because getting it wrong is a common first-install failure.

Frigate requirements at a glance

Frigate’s installation and recommended hardware pages set a few more requirements that change what you buy. Both were read on 2026-10-10.

Sizing the storage is a separate job from picking the box: size it on recorded bitrate, not camera megapixels.

The camera-count ladder

Frigate publishes activity-based guidance in its planning docs, and names specific machines in its recommended hardware list. Neither is a buying table on its own. Combined, they are:

Frigate’s own capability notes for the named machines: the Beelink EQ13 (N100) “can run object detection on several 1080p cameras with low-medium activity”; the 1220P class “can handle a large number of 1080p cameras with high activity”; the 125H class “can handle a significant number of 1080p cameras with high activity.”

Activity matters as much as count. Frigate gates object detection behind motion detection, so a quiet scene costs far less to run than a busy one. Six cameras pointed at a quiet side yard is a genuinely different load from six cameras on a busy street. Frigate’s own tiering is written in those terms, which is why a flat “cameras → CPU” table from any other site is misleading. If your cameras see constant motion, size up one row.

Best GPU for Frigate: iGPU, Arc or Nvidia

A GPU does two separate jobs in Frigate, and the best card for one is not always the best for the other.

Decoding: Frigate’s video decoding docs say it can use most Intel integrated GPUs and Arc GPUs, modern AMD integrated and discrete GPUs, and most modern Nvidia GPUs to decode camera streams.

Detection: each vendor has its own route.

• Intel: the OpenVINO detector runs on 6th-generation and newer Intel iGPUs, Intel Arc A-series and B-series cards, and Intel NPUs.
• Nvidia: the ONNX detector uses a GPU that supports the CUDA 12.x libraries. The host driver must be version 545 or newer, and the GPU needs Compute Capability 5.0 or higher, which Frigate says generally means a Maxwell-era card or newer. The host also needs the Nvidia container runtime.
• AMD: the ROCm detector runs on many discrete AMD GPUs. Frigate says it supports limited model architectures and runs best on discrete cards.

Frigate’s recommended-hardware page publishes inference times for some of these parts, using the YOLOv9 model at 320 pixels:

Source: Frigate’s recommended hardware page, read 2026-10-10. Frigate notes that speeds vary greatly with the GPU and the model.

The table measures inference time and nothing else. It says nothing about price or power draw, which is why the iGPU stays the default pick in the ladder above. A discrete GPU makes sense when you already own one, when a busy 12-camera install outgrows an iGPU, or when you want larger models. Frigate’s planning guide names a discrete GPU as one option at that scale.

The Coral question, settled by the people who make Frigate

This is where most of the internet is out of date, including the AI Overview’s own sources.

Frigate’s hardware documentation now states plainly: “The Coral is no longer recommended for new Frigate installations, except in deployments with particularly low power requirements or hardware incapable of utilizing alternative AI accelerators for object detection.” Frigate adds that it “will continue to provide support for the Coral TPU for as long as practicably possible,” so nobody with a working Coral is stranded.

Two things follow, and they save real money:

On an Intel box, the free option is the right one. The integrated GPU handles decoding via Quick Sync and runs detection through Frigate’s OpenVINO detector. One chip, both jobs, zero add-in cost. For the 1–12 camera range that covers most homes, this is the whole answer.

Where a dedicated accelerator genuinely helps, it’s Hailo now. Frigate states that it “supports both the Hailo-8 and Hailo-8L AI Acceleration Modules on compatible hardware platforms,” and lists them among its officially supported detectors.

I may earn a commission if you buy through these affiliate links. As an Amazon Associate I earn from qualifying purchases.

The document status matters more than the old listing order:

On an Intel machine you buy neither. OpenVINO on the iGPU is already paid for. Reach for a Hailo when you are on hardware with no usable iGPU, such as a Raspberry Pi 5, or when you are past a dozen busy cameras.

One shopping warning. An automated Amazon US product search for “Hailo-8L M.2” on 2026-07-30 returned accessories, a Raspberry Pi case, an NVMe adapter HAT and unrelated paid placements alongside accelerator modules. Read the title on every listing and match the exact module before ordering. The stored listing figures from that lookup are no longer published.

Boxes worth buying

One caution on the N150 row. Frigate’s recommended-hardware page, read 2026-10-10, says the Beelink EQ14 has known compatibility issues and should be avoided for now. Check the exact model name on any N150 Beelink listing against that warning before you order.

Account for what a barebone box omits. A barebone mini PC ships without RAM or an SSD, so compare its fully configured total with a complete box rather than comparing the two listing headlines. Use current offers for the required SO-DIMM kit and NVMe drive before deciding. We work through the configuration trade-offs in our best mini PCs for a homelab guide.

If your NVR will also be your smart-home hub, the sizing question changes and we cover it separately in best mini PC for Home Assistant. That page is about a Home Assistant box that also watches a few cameras. This page is about a machine whose main job is cameras. If you want both roles isolated on one machine, our Home Assistant on Proxmox guide covers the virtualisation route.

Storage: where the real money leaks

Frigate’s docs are clear on roles. SSDs “are designed with advanced wear-leveling algorithms” and give you a responsive UI and quick clip retrieval, which suits the OS and database. For the relentless 24/7 write load of recording, Frigate points at NVR-rated drives like WD Purple and Seagate SkyHawk, “engineered for 24/7 operation and continuous write workloads.”

The 2026-07-30 retailer lookup established the exact models below. Its stored listing prices and derived cost-per-TB figures have been removed; use the paired Amazon and Newegg controls to compare current offers.

Run the current cost-per-TB calculation before buying a small drive, and check both brands at your target capacity every time. Model numbers, capacities and the TOPS figures above are as given in the retailer listings on 2026-07-30, not verified against manufacturer spec sheets. On annual workload rating, Seagate states that SkyHawk drives “can support workloads up to 180TB per year,” the same figure the WD Purple listing quotes. On that specification the two are equivalent. We have not tested any of these drives; every rating here is the manufacturer’s own. For drive selection beyond surveillance duty, see best NAS hard drives.

Size it before you buy it. Retention days, camera count, resolution, and codec drive the number far more than intuition suggests. Going from 1080p to 4K on eight cameras is not a small adjustment. Put your actual numbers into the Frigate NVR storage calculator and buy against the result. It also computes the /dev/shm value Frigate needs, which is the setting most first-time installs get wrong.

Don’t forget the network side

Cameras need power and bandwidth, and PoE switches are budgeted in watts, not ports. A switch with eight PoE ports and a 60W budget will not run eight 15W cameras. Work out the real figure with our PoE budget calculator before ordering, because “8-port PoE” tells you almost nothing about what it can actually power.

What we’d actually buy

• Four 1080p cameras on a quiet property: Beelink S12 Pro (N100, 16GB), OpenVINO on the iGPU, and one 4TB SkyHawk. No accelerator.
• Eight cameras, mixed activity: the same 16GB tier is Frigate’s “highly recommended” line. Stay on it, step the CPU to a Core i-series box, and move to 6TB or dual 4TB drives.
• Twelve or more, busy scenes: this is where Frigate’s docs say multiple accelerators or a real GPU, and where a Hailo-8 becomes relevant.

The summary: for the majority of home installs, Frigate is not demanding hardware. It asks for two specific things, AVX2 and hardware decoding, and is generous about everything else. Most people over-buy the accelerator and under-buy the drive. Do it the other way around.

Sources

Frigate’s own documentation, fetched 2026-07-30, with the recommended-hardware page, the installation page and the video-decoding page re-read 2026-10-10:

• Recommended hardware — named machines, capability wording, Coral and Hailo status
• Planning a new installation — RAM tiers, AVX requirement, camera-count guidance, drive recommendations
• Video decoding — hardware acceleration guidance
• Object detectors — OpenVINO and Hailo detector configuration
• Installation — host, shared-memory and Docker requirements, read 2026-10-10

An automated Amazon US retail lookup was completed on 2026-07-30, with product titles read individually to exclude accessories and bundles. Its stored listing prices and derived per-TB figures are no longer published; the paired Amazon and Newegg controls lead to current offers. All capability claims are Frigate’s own words, quoted and linked.