How Frigate uses a mini PC
Frigate does two different jobs. First it decodes every camera stream so it can look for motion; that work is done by the CPU, or much more efficiently by the iGPU through hardware acceleration. Then it sends frames with motion to a detector, which runs the AI model that decides whether you are looking at a person, a car or a swaying branch. Frigate's docs are explicit that a Coral "does not help with decoding video streams", and they call hardware-accelerated decoding on an integrated or discrete GPU highly recommended. A used Intel mini PC handles both jobs on the same chip.
OpenVINO on the Intel iGPU
Frigate's OpenVINO detector runs on 6th-gen Intel (Skylake) and newer platforms with an iGPU, on Intel Arc GPUs, and on Intel NPUs; it also runs on most modern AMD CPUs, though Intel doesn't officially support that. The hardware page lists example inference times with MobileNetV2: about 15–25 ms on Intel HD 620, about 15 ms on an N100, and about 10 ms on UHD 730 and Iris Xe.
Frigate's rule of thumb for detector capacity is 1000 divided by the inference time. Its example uses a Coral: at 10 ms, it tops out at 100 detections per second. The same arithmetic puts an N100 at roughly 66 per second. Detection only runs on frames with motion, so that budget stretches across more cameras than the raw number suggests. If you approach the limit, Frigate's docs say to tune motion masks first.
What about a Google Coral?
The Coral was the default Frigate accelerator for years, but the hardware page now says it is "no longer recommended for new Frigate installations, except in deployments with particularly low power requirements or hardware incapable of utilizing alternative AI accelerators." Frigate will keep supporting it because it is still one of the most power-efficient detectors. The USB version needs no host driver but lacks the automatic throttling of the other versions; the M.2 and PCIe versions need a driver installed on the host. Hailo-8 and Hailo-8L modules are now near the top of the docs' detector list. On a used Intel mini PC, start with OpenVINO on the iGPU and add an accelerator only if inference times climb.
Hardware video decoding: VAAPI or QSV
For Intel graphics, Frigate's docs recommend the preset-vaapi hardware-acceleration preset for generations 1 through 12, because VAAPI picks the right profile automatically for both H.264 and H.265 streams. For 13th gen and newer, and for Arc GPUs, use the preset-intel-qsv-h264 or preset-intel-qsv-h265 presets to match your cameras. Some older Intel CPUs need LIBVA_DRIVER_NAME=i965 because the default iHD driver doesn't support them. If you also run a media server, it uses the same Quick Sync hardware; the Quick Sync by generation guide shows what each generation decodes.
Camera choice matters as much as the PC. Frigate recommends cameras that output H.264 video and AAC audio and that offer multiple substreams, so a low-resolution stream can feed detection while the main stream is recorded. Its docs warn against Wi-Fi cameras, and suggest sticking to 5 MP and lower if you use Reolink.
How many cameras?
Frigate doesn't publish a per-CPU camera count, because resolution, frame rate and how much movement each scene has all change the load. Its recommended-hardware table is the best guide: the Beelink EQ13 (N100) can run object detection on several 1080p cameras with low to medium activity, while higher tiers built on Core-class Intel mobile chips (listed as Intel 1120p and Intel 125H) handle large numbers of 1080p cameras with high activity, the 125H adding an NPU. Higher resolutions and frame rates mean more decoding work, so set the detect stream to a modest resolution on the camera itself.
Storage for recordings
Recordings are the part that grows. Keep the OS and Frigate's database on the NVMe drive and give recordings their own disk. On 1-liter business tinies such as the M720q, the 2.5-inch SATA bay is the natural home for a large HDD or SSD. The Beelink EQ13 gives up the 2.5-inch bay its EQ12 predecessor had in exchange for a second M.2 NVMe socket, which works just as well for a high-capacity recording drive. You can also write to a NAS share, as long as the network path is reliable.
Running Frigate alongside Home Assistant
Most people pair Frigate with Home Assistant, and both fit on one mini PC. If you run them under Proxmox, decide early how the Frigate container or VM will reach the iGPU's /dev/dri device, since that is what both decoding and OpenVINO depend on. Put cameras on their own VLAN or second NIC and block them from the internet; network services covers the router side. If your plans include an AI agent server or workflow automation on the same box, size RAM for all of it. For the N100 family, see N100 vs N150 vs N305.