Hardware Recommendation for 12-Camera Frigate Installation

[arnobfactorynext]

Hi everyone, I’m planning a 12-camera Hikvision IP camera setup with Frigate and Home Assistant for real-time person detection/occupancy and lighting automation.
Cameras will be wired/PoE, with the main stream for streaming and substream (~10 FPS) for detection.
I’m planning to run Frigate 24/7 and would like some headroom for future expansion.
What CPU, RAM, GPU/iGPU or may be a miniPC would you recommend for this workload?
Also, what are the other things would be appropriate for 12 cameras?
I’d especially appreciate recommendations from anyone currently running 10–15+ cameras with Frigate.

[inkeep-ai-assistant[bot]]

For 12 cameras, you're in the moderate-to-high simultaneous activity range. Setups with 6-12 cameras, especially where several capture activity at the same time, increase the load on your object detection hardware and require more processing power. 12+ cameras or highly active zones will likely need multiple entry-level AI accelerators or a more powerful single unit such as a discrete GPU.(1)

CPU and platform

Frigate requires a CPU with AVX + AVX2 instructions. Most modern CPUs (post-2011) support these, but they are generally absent in low-power or budget processors — Intel Celeron and Pentium models prior to the 2020 Tiger Lake generation typically lack AVX.(1)

For mini PCs, the documented options that suit your scale:

• Intel 1120p — can handle a large number of 1080p cameras with high activity.(2)
• Intel 125H — can handle a significant number of 1080p cameras with high activity, and includes an NPU for more efficient detection in 0.17+.(2)
• Beelink EQ13 (N100) — can run object detection on several 1080p cameras with low-medium activity, with dual gigabit NICs for an isolated camera network. This is likely under-spec for 12 active cameras.(2)

Note the Beelink EQ14 has known compatibility issues and should be avoided.(2) Many of these mini PCs ship with Windows and will need Linux installed.(2)

Detection hardware

Use of a GPU or AI accelerator is highly recommended; AI accelerators outperform even the best CPUs with very little overhead.(3) Options that fit:

• OpenVINO on an Intel iGPU (6th gen or newer), Intel Arc GPU, or Intel NPU. Documented inference times include Intel N100 at ~15 ms for MobileNetV2 (can only run one detector instance), Intel NPU at ~6 ms, Intel Arc A310 at ~5 ms, and Intel Iris XE at ~10 ms.(2)
• Nvidia GPU via the onnx detector, requiring driver >=545 and Compute Capability of 5.0 or greater, plus nvidia-container-runtime on the host.(2)
• Hailo-8/8L M.2 modules, defaulting to YOLOv6n (~7 ms on Hailo-8, ~11 ms on Hailo-8L).(2)

The Coral is no longer recommended for new installations except where power requirements are particularly low or hardware can't use alternatives.(2)

You can size your detector from inference speed: with an inference speed of 10ms, your detector handles roughly 100 detections per second (1000ms / 10ms = 100).(4) Note that openvino and onnx support multiple detector instances to share the workload when a single detector can't keep up — useful with GPUs that have enough VRAM.(4)

Video decoding

Modern CPUs with integrated GPUs (Intel Quick Sync, AMD VCN) or dedicated GPUs significantly offload video decoding from the main CPU, which is highly recommended for multiple cameras.(1) Video decoding is one of the most CPU-intensive tasks in Frigate, and an AI accelerator does not assist with decoding.(4) Hardware acceleration lets you support 2-3x more cameras on the same hardware.(4) Decode burden grows exponentially with resolution and frame rate — a 4K 30 FPS stream requires roughly 4x the processing of 1080p at the same rate.(4)

Your stream/FPS plan

Reconsider the ~10 FPS substream. detect.fps controls how many times per second Frigate runs object detection and does not need to match your camera's frame rate — the default of 5 is correct for the vast majority of cameras.(5) Running at 10 doubles the frames eligible for detection and is one of the largest contributors to detector CPU usage.(4) Most users who raise it above the default don't need to.(5)

Also tune the substream resolution: the ideal detect resolution is one where objects fit inside the model dimensions (320x320), and higher resolutions do not improve accuracy because detail is lost in the resize.(5) Don't match detect resolution to model resolution — the detect stream should match your camera's aspect ratio, for example 1280x720.(4) Keep detection and recording streams on the same aspect ratio for the best UI experience.(5)

Recording stream should be the resolution you wish to store, typically the highest your camera supports, set to 15 fps in camera firmware.(5)

Cameras

Hikvision is a recommended brand — models have multiple streams with configurable resolutions and frame rates and rock solid streams, with large-sensor models known for excellent night image quality.(2) Cameras outputting H.264 video and AAC audio offer the most compatibility with all features of Frigate and Home Assistant.(5)

Storage

Use an IP camera storage calculator such as IPConfigure Storage Calculator to estimate disk space from your camera settings.(1) SSDs offer high speed and responsiveness, handle the many small writes of continuous recording well, and modern drives' wear concerns are largely no longer valid.(1) HDDs offer the best cost per gigabyte for long-term archives — consider NVR-rated drives such as Western Digital Purple or Seagate SkyHawk, engineered for 24/7 operation and continuous write workloads.(1)

Network

Dual NICs let you set up a dedicated private network for your cameras where they can be blocked from accessing the internet.(2) Stick with wired/PoE as you planned — WiFi cameras are not recommended since their streams are less reliable and cause connection loss and/or lost video data.(2)

RAM is not specified in the documentation for this scale, so I can't give you a grounded number there.

📚 Sources:

1. Planning a New Installation
2. Recommended hardware
3. Frigate NVR™ - Realtime Object Detection for IP Cameras
4. High CPU Usage
5. Camera setup

---

Have another question? Just tag @inkeep.

[PaulOckenden]

Are you using it with a coral?
No need for my use, the Intel GPU copes just fine. (5x 8MP cameras)

If I had l loads more cameras I'd probably add a Hailo8.

But my main point is that an EQ14 runs Frigate (Docker install on Debian) just fine.

[NickM-27]

Right, the compatibility note is related to it not working with the coral

[PaulOckenden]

Right, the compatibility note is related to it not working with the coral

See #18850 - plenty of folks using an EQ14 with Coral. The person who was having problems near the top of the page got it working just fine if you read further down.

[arnobfactorynext]

Right, the compatibility note is related to it not working with the coral

See #18850 - plenty of folks using an EQ14 with Coral. The person who was having problems near the top of the page got it working just fine if you read further down.

how many camera are you using with your current system? and how is the detection speed?

[arnobfactorynext]

Are you using it with a coral?

Noh, for now i m just using it in Home Assistant os with frigate full access add ons in a minipc, qbits lania mini f

[sabraaz]

I run 64 cameras using 4 x hailo8's and an RTX 2000 ADA 16gb for decoding. It uses about 9gb of gpu memory, 29gb of system memory, and 6-7 cores of a Xeon 6208U cpu which is a couple years old at 100%. I'm using the yolo9s 320x320 model with most cameras running at 1280x720 @ 5fps per recommendations.

With 12 cameras you could most likely get away with using a modern intel cpu that comes with the gpu/npu. The caveat with any setup is the amount of detections more than the number of cameras. The busier the scene you monitor, the more processing you'll need. Most of my scenes are low activity and outdoor. Even with 400 fps of detection power available, it can be overwhelmed in partially shaded areas if clouds obstruct the sun then unobstruct the sun rapidly. That generates areas of motion that will receive object detection scanning. Combine that with wind and it's a short lived nightmare that you can't do much about.

[arnobfactorynext]

I run 64 cameras using 4 x hailo8's and an RTX 2000 ADA 16gb for decoding. It uses about 9gb of gpu memory, 29gb of system memory, and 6-7 cores of a Xeon 6208U cpu which is a couple years old at 100%. I'm using the yolo9s 320x320 model with most cameras running at 1280x720 @ 5fps per recommendations.

With 12 cameras you could most likely get away with using a modern intel cpu that comes with the gpu/npu. The caveat with any setup is the amount of detections more than the number of cameras. The busier the scene you monitor, the more processing you'll need. Most of my scenes are low activity and outdoor. Even with 400 fps of detection power available, it can be overwhelmed in partially shaded areas if clouds obstruct the sun then unobstruct the sun rapidly. That generates areas of motion that will receive object detection scanning. Combine that with wind and it's a short lived nightmare that you can't do much about.

Can we talk? can i have your mail or anything else?

[arnobfactorynext]

I run 64 cameras using 4 x hailo8's and an RTX 2000 ADA 16gb for decoding. It uses about 9gb of gpu memory, 29gb of system memory, and 6-7 cores of a Xeon 6208U cpu which is a couple years old at 100%. I'm using the yolo9s 320x320 model with most cameras running at 1280x720 @ 5fps per recommendations.

With 12 cameras you could most likely get away with using a modern intel cpu that comes with the gpu/npu. The caveat with any setup is the amount of detections more than the number of cameras. The busier the scene you monitor, the more processing you'll need. Most of my scenes are low activity and outdoor. Even with 400 fps of detection power available, it can be overwhelmed in partially shaded areas if clouds obstruct the sun then unobstruct the sun rapidly. That generates areas of motion that will receive object detection scanning. Combine that with wind and it's a short lived nightmare that you can't do much about.

Thanks, this is very helpful. My cameras will mostly monitor indoor offices/rooms for person occupancy and lighting automation, so activity should be relatively low compared with outdoor scenes. I plan to use 12 Hikvision cameras at 10 FPS detection, with 1280×720-ish substreams. For this workload, what CPU/platform would you recommend for a mini PC, and how much RAM would you suggest?

[meeki007]

Good Day @arnobfactorynext
I run a 108 camera system for a warehouse and office. been running frigate from ver 0.14 (early 2024) to now (late 2026)

"What CPU, RAM, GPU/iGPU or may be a miniPC would you recommend for this workload?"

I'm in the process of migration to using the OrangePi 5 Max as it has a proper m.2 2280 slot for mounting a SSD for storage. The built in NPU 6TOPS, Ram 16GB LPDDR5, and the Built in Arm Mali-G610 GPU.

You list 1280×720 detection and that will lower the CPU usage on detect..
I'm running most cameras at 3k to 4k record and detect because they are old skool cameras that have BS substreams of 640x480
even recording and detect with: detect: width: 2592 height: 1520 I'm able to run 8 Cameras with 1/2 the OrangePi 5 Max hardware resources.
I don't know if the unit will handle 12 cameras yet as storage space requirements are the issue. I'm using 8TB SSD m.2 Also the system resource usage below is with semantic_search model: "jinav1" and snapshots enabled.

.

Also I run the cameras and record at 15FPS and I save allot of system resources by config-ing the Object detect.

detect:
  # Optional: enables detection for the camera (default: shown below)
  enabled: true
  # Optional: desired fps for your camera for the input with the detect role (default: shown below)
  # NOTE: Recommended value of 5. Ideally, try and reduce your FPS on the camera.
  fps: 3
  # Optional: Number of consecutive detection hits required for an object to be initialized in the tracker. (default: 1/2 the frame rate)
  min_initialized: 3
  # Optional: Number of frames without a detection before Frigate considers an object to be gone. (default: 5x the frame rate)
  max_disappeared: 15
  # Optional: Configuration for stationary object tracking
  stationary:
    # Optional: Stationary classifier that uses visual characteristics to determine if an object
    # is stationary even if the box changes enough to be considered motion (default: shown below).
    classifier: true
    # Optional: Frequency for confirming stationary objects (default: same as threshold)
    # When set to 1, object detection will run to confirm the object still exists on every frame.
    # If set to 10, object detection will run to confirm the object still exists on every 10th frame.
    interval: 9
    # Optional: Number of frames without a position change for an object to be considered stationary (default: 10x the frame rate or 10s)
    threshold: 30

If you do go the OrangePi 5 Max route it has the same mounting holes as a Raspberry Pi I use this to rack mount them. https://www.amazon.com/dp/B0F6NMNJQ7?ref=fed_asin_title

I hope all this helps you make a selection but It mostly comes down to TRY and FIND-OUT

Oh and if you have to expand and don't want to play what NVR is the camera on Check out my post on using Node-red and world maps to create a camera map that links you to the right system
#14858 (comment)

[Cold-Lemonade]

I have a 14-camera system which I run on a Beelink EQi12 with an Intel i3-1220P cpu and Hailo-8. Even with the Hailo-8 and motion masks, the system will skip motion detected frames if there is too much activity. This is a capacity limit related to the Hailo-8 I believe. To address that, I limit the object detection to 10 cameras and focus on just finding "person". For storage, I use a Terra-Master D4-320 with 64 TB LVM pool split into two 32 TB mirrors. I was lucky to buy used HDDs before prices went crazy. Also, five cameras are 2 MP, one is 8 MP, the rest are 4 MP. CPU usage is around 20% and iGPU is 4% or less. My storage gives me about 30 days of 24/7 recordings.

I am not sure if the following gets enough mention here. You should isolate your cameras on a separate network. The Beelink EQi12 is nice for this because it has two nics. One is on my internet facing network, the other the network that only connects to the cameras. Nearly all, if not all, Chinese-made cameras will call home and share God-knows-what with them. Isolate the cameras so that can't call home.