Recommended Server Specs for ~50 Camera Setup (Frigate / AI Detection)

[dkeller-collab]

Hi everyone,

I’d like to open a discussion regarding optimal server specifications for handling approximately 50 IP cameras with AI-based detection (e.g., Frigate or similar NVR systems).

🎯 Use Case
~50 cameras (mix of 1080p / possibly some 4MP)
Continuous recording + object detection
Real-time or near real-time inference
Likely using Frigate with a GPU or hardware acceleration

💻 Proposed Hardware (Initial Idea)
CPU: Intel i7-13700K or i9-13900 (with iGPU for Quick Sync)
GPU: NVIDIA RTX 4080
RAM: 64GB DDR5
Storage:
2TB NVMe (OS + cache)
10TB HDD (recordings)
PSU 850W–1000W Gold
NZXT H7 Flow RGB Mid-Tower ATX Airflow Case with RGB Fans

❓ Key Questions
Is a high-end GPU like an RTX 4080 overkill for ~50 cameras, or justified?
Would relying more on Intel Quick Sync (iGPU) be more efficient than a discrete GPU?
What models (YOLOv8/YOLOv9/Frigate default) are best suited for this scale?
Any bottlenecks I should anticipate (CPU vs GPU vs disk I/O)?
Would a multi-server setup be more stable than a single powerful machine?

⚙️ Additional Considerations
Potential use of hardware acceleration (VAAPI / NVDEC)
Balancing detection FPS vs number of streams
Long-term scalability (future expansion beyond 50 cameras)

🙌 Looking for Feedback
I’d really appreciate insights from anyone running similar-scale deployments:

Real-world performance benchmarks
Recommended optimizations
Lessons learned (especially around stability and scaling)

Thanks in advance!

[inkeep-ai-assistant[bot]]

For a ~50 camera deployment, here's what the documentation indicates:

Scale Classification

Commercial installations with 40+ cameras require robust object detection capabilities and will likely need a modern discrete GPU. (1)

GPU Considerations

Is an RTX 4080 overkill?

For commercial installations, it's important to verify the number of supported concurrent streams on your GPU—many consumer GPUs max out at ~20 concurrent camera streams. (1)

This is a critical consideration for your 50-camera setup. The RTX 4080 would handle object detection efficiently, but you need to verify its concurrent stream limits for decoding.

Intel Quick Sync vs Discrete GPU

Modern CPUs with integrated GPUs (Intel Quick Sync) can significantly offload video decoding from the main CPU, freeing up resources. This is highly recommended, especially for multiple cameras. (1)

For your setup, using both would be ideal—Intel Quick Sync for video decoding and the Nvidia GPU for inference.

Detector Performance

With the -tensorrt Frigate image, Nvidia GPUs are automatically detected and used with the ONNX detector. (2)

Known inference times on similar Nvidia hardware: (3)

• RTX 3050: YOLOv9 t-320: 8ms, s-320: 10ms
• RTX 3070: YOLOv9 t-320: 6ms, s-320: 8ms

YOLOv9 supports CUDA Graphs for optimal Nvidia performance. (2)

Model Recommendations

For Nvidia GPUs, these models are supported with CUDA Graph acceleration: (2)

• YOLOv9 ✅
• RF-DETR ✅
• YOLOX ✅

YOLO-NAS and D-FINE are not supported by CUDA Graphs and may have reduced performance. (2)

Multiple Detectors

When using many cameras, one detector may not be enough to keep up. Multiple detectors can be defined assuming GPU resources are available:

detectors:
  onnx_0:
    type: onnx
  onnx_1:
    type: onnx

(2)

Potential Bottlenecks

1. Concurrent stream limits on consumer GPUs for decoding (1)
2. Simultaneous activity - with 50 cameras and high activity zones, detection demand increases significantly (1)

Multi-Server Consideration

The documentation doesn't explicitly address multi-server vs single-server stability for this scale. However, given the concurrent stream limitations noted for consumer GPUs, splitting the workload could help avoid hitting those limits.

Hardware Acceleration Config

For your setup, you'd configure hardware acceleration in ffmpeg settings. The default hwaccel_args: "auto" will attempt to auto-detect your hardware. (4)

---

Bottom line: Your proposed hardware is reasonable for this scale. The RTX 4080 isn't overkill given the 50-camera count and need for real-time inference. However, verify concurrent decode stream limits, and consider running multiple detector instances to handle the load.

📚 Sources:

1. Planning a New Installation
2. Supported Hardware
3. Recommended hardware
4. Full Reference Config

---

Have another question? Just tag @inkeep.

[easy-and-simple]

No mater whether 1 computer can handle 50 cameras, as in any way It is unprofessional to set 50 cameras on 1 computer, because if it crashes you will get 50 not working cameras and will be fully blind.
Professional way is to split cameras on few computers in different locations in that way, so if some computer crash to have at least 1 camera cowering every area. Interface is remote so you can watch these few computers from 1 place