Running meaningful AI analytics on security camera feeds — real-time object detection, people counting, ALPR, behavioral analysis — requires compute infrastructure that scales with camera count and analytics complexity. This guide covers what hardware is actually required for self-hosted AI surveillance, how to size it, and when cloud AI becomes the lower-cost and lower-complexity alternative.
Why AI Surveillance Requires Dedicated Hardware
Motion detection — comparing pixel changes between frames — runs adequately on CPU. AI object detection is different: it runs neural network inference on each processed frame, applying a trained model to classify objects in the image. Modern detection models (YOLOv8, EfficientDet, MobileNet SSD) are computationally intensive even at reduced input resolution.
The key constraint: inference must complete faster than the frame rate being processed. For real-time detection at 10 FPS across 4 cameras, the inference hardware must process 40 frames per second. At 30 FPS across 10 cameras, that’s 300 FPS. CPU-only inference cannot approach these throughputs — GPU acceleration or dedicated AI hardware is required for any meaningful deployment.
Option 1: Google Coral TPU (Best for Low-Power Small Deployments)
Coral TPU Specs
- Edge TPU ASIC — 4 TOPS
- USB or M.2/PCIe form factor
- USB 3.0: $60–$80
- M.2 / Mini PCIe: $40–$50
- Dual Edge TPU: $100–$120
- Power draw: 2–4W
- Inference: 50–80 FPS (TFLite)
Best for: Small residential or micro-commercial deployments of 1–8 cameras where power consumption and hardware cost are primary constraints. The Coral TPU is the primary hardware recommendation for Frigate NVR deployments at home scale.
Limitations: Only runs TensorFlow Lite models compiled specifically for the Edge TPU. Model customization is limited. For cameras above 4MP or complex scenes requiring larger models, the Coral begins to bottleneck. The USB variant has thermal throttling limitations under sustained load.
Availability note: Coral Edge TPU products have experienced significant supply constraints since 2022 due to semiconductor shortages and Google’s reduced focus on the product line. Pricing has risen and availability is inconsistent through authorized channels.
Option 2: Intel OpenVINO / Integrated GPU (CPU Server Deployments)
Intel iGPU / OpenVINO
- Integrated Intel GPU
- No discrete GPU required
- Intel NUC: $200–$600
- Power: 15–45W (full system)
- Inference: 20–60 FPS (OpenVINO)
- Best model: YOLOv8n, SSD MobileNet
Best for: 4–12 camera deployments on energy-efficient Intel NUC or similar mini-PC servers. Frigate supports Intel QuickSync/OpenVINO acceleration natively. This is the most cost-effective option for deployments that outgrow a Coral TPU.
Practical example: An Intel NUC 12 or 13 with an Intel Core i5/i7 and its integrated GPU can handle 8–12 cameras at moderate analytics density using Frigate with OpenVINO — running at sub-30W for the entire system.
Option 3: NVIDIA GPU (Mid to Large Deployments)
For deployments above 15 cameras, or deployments requiring complex AI models (larger YOLO variants, multi-class detection, behavioral analytics), a discrete NVIDIA GPU running CUDA-accelerated inference is the appropriate hardware class.
| GPU Model | VRAM | Estimated Cameras | Current Price | Power (TDP) | Best Use |
|---|---|---|---|---|---|
| RTX 3060 | 12GB | 10–20 | $280–$380 | 170W | Small commercial, 15–25 cameras |
| RTX 3060 Ti | 8GB | 15–25 | $320–$420 | 200W | Mid-size commercial |
| RTX 4070 | 12GB | 20–40 | $480–$580 | 200W | Medium enterprise |
| RTX 4080 | 16GB | 35–60 | $900–$1,100 | 320W | Large enterprise, single site |
| RTX 4090 | 24GB | 50–80 | $1,600–$2,000 | 450W | Large enterprise, multiple streams |
| NVIDIA A2000 (Pro) | 12GB | 20–35 | $600–$900 | 70W | Server rack, low TDP, 24/7 duty cycle |
| NVIDIA A4000 (Pro) | 16GB | 35–70 | $900–$1,400 | 140W | Enterprise server, high reliability |
Important: Consumer GPUs and 24/7 Duty Cycle
Consumer GPUs (RTX series) are rated for consumer workloads, not 24/7 inference. Running an RTX 4070 at full inference load continuously may reduce its lifespan compared to rated specs. For production deployments requiring high reliability, professional GPUs (NVIDIA A-series) designed for continuous compute workloads are the appropriate choice — despite their higher cost per TFLOPS.
Total Infrastructure Cost: Beyond the GPU
GPU cost is one line item. The full AI inference server for a commercial self-hosted deployment includes:
| Component | Cost Range | Notes |
|---|---|---|
| GPU (RTX 4070 example) | $480–$580 | For 20–40 cameras |
| Server chassis / workstation | $400–$1,200 | Must accommodate full-length GPU + sufficient PCIe lanes |
| RAM | $80–$200 | 32–64GB recommended for multi-camera VMS + AI workload |
| Storage (SSD for OS + footage) | $100–$600 | NVMe SSD for OS; HDD array for footage retention |
| UPS (uninterruptible power supply) | $150–$500 | Required to prevent corrupt drives from power loss |
| Power cost (3yr @$0.12/kWh) | $380–$1,200 | 200–400W system running 24/7 × 3 years |
| IT setup labor | $1,500–$6,000 | 20–80 hrs × $75/hr — OS, CUDA, VMS, AI pipeline configuration |
| Total estimated range | $3,100–$10,300 | Before ongoing maintenance labor |
Cloud AI vs. On-Premise GPU: When Each Makes Sense
| Scenario | On-Premise GPU | Cloud AI (iFovea) |
|---|---|---|
| 1–4 cameras, home/small business | Coral TPU is cost-effective | Cloud subscription may cost more over 3 years |
| 5–20 cameras, single site | RTX 3060 / Intel iGPU viable | Cloud often competitive on TCO |
| 20+ cameras, multi-site | GPU per site — multiplies costs | Cloud scales without additional hardware |
| Air-gapped environment | Required (no internet = no cloud AI) | Not applicable |
| ALPR, people counting, forensic search | Requires high-end GPU + software integration | Native — included in subscription |
| Limited IT/ops team | High ongoing maintenance burden | Platform managed — minimal operator burden |
Want to Compare Cloud AI vs. On-Premise GPU for Your Deployment?
Share your camera count and analytics requirements — we’ll build a side-by-side cost model for your specific scenario.
