System requirements
Hardware
Nvidia
Agent Configuration: 0.1 CPU cores, 0.4 GB RAM.
Per-thread consumption: 20–25 FPS for FP32 computation accuracy. All CPU consumption is calculated as the number of threads with hyperthreading.
| Name | CPU | RAM | VRAM | GPU (FP32) |
|---|---|---|---|---|
| Video Decompression | 0.2 threads (0.15 cores) | – | – | – |
| Face Identification | 1.1 threads (0.7 cores) | 1.3 GB+3 GB per 1 million faces | 0.25 Gb | 1.2 TFLOPs |
| Face Identification (CPU only) | 2.3 threads (1.6 cores) | 3.5 Gb+3 Gb for 1 million faces | – | – |
| Body Detection and Skeletal Tracking | 1.3 threads (0.9 cores) | 0.4 Gb | 0.3 Gb Cuda 0.2 Gb Tensor | 1.4 TFLOPs Cuda 1.0 TFLOP Tensor |
| Human Activity Recognition (HAR) is added to the body detector consumption. | 0.2 threads (0.13 cores) | 0.2 GB | 0.3 GB | 3.0 TFLOP Cuda 2.5 TFLOP Tensor |
| UAD new objects (700×700) | 1.9 threads (1.3 cores) | 1.9 GB | 1.8 GB | 3.8 TFLOP |
| Object Detection (phone, cigarettes, etc.) | 0.4 threads (0.3 cores) | 0.17 GB | 0.15 GB | 1.2 TFLOPs Cuda 0.8 TFLOPs Tensor |
| Driver/Operator Monitoring (DSM) | 1.0 threads (0.7 cores) | 1.0 GB | 1.4 GB | 3.1 TFLOPs |
Desktop Test Machine Specifications (Nvidia)
- OS: Ubuntu 20.04
- CPU: AMD Ryzen 9 5950X @ 3.4Ghz: 16 cores (32 threads)
- RAM: 118GB DDR4
- VRAM: 11GB
- Disk: 500GB Samsung SSD 870 (read 560 MB/s, write 530 MB/s)
- GPU: NVIDIA GeForce GTX 1080 Ti 10.6 TFLOPS
NVIDIA GPU Comparison by FP32/FP16 Performance
| Graphics Card | VRAM Size | FP32 (single precision) | FP16 (half precision) |
|---|---|---|---|
| GTX 1080 Ti | 11 GB GDDR5X | 11 TFLOPS | 0.2 TFLOPS |
| RTX 2080 Ti | 11 GB GDDR6 | 13 TFLOPS | 26 TFLOPS |
| RTX 3060 Ti | 8 GB GDDR6 | 16 TFLOPS | 32 TFLOPS |
| RTX 3080 Ti | 12 GB GDDR6X | 34 TFLOPS | 68 TFLOPS |
| RTX 4070 | 12 GB GDDR6X | 22 TFLOPS | 58 TFLOPS |
| RTX 4090 | 24 GB GDDR6X | 82 TFLOPS | 165 TFLOPS |
Edge (Sophgo)
Agent Configuration: CPU 0.1 cores, RAM 0.4 GB
Thread Consumption – 20–25 FPS for FP16 computation accuracy.
| Name | CPU | RAM | RAM NPU | NPU (FP16) |
|---|---|---|---|---|
| Video Decompression | 0.3 Cores | 0.05 GB | 0.17 GB | – |
| Passenger Counting (APC) | 0.6 Cores | 0.05 GB | 0.20 GB | 1.5 TOPS |
| Driver Monitoring (DSM) | 1.3 cores | 0.10 GB | 0.25 GB | 2.3 TOPS |
| Body Detection and Skeletal Tracking | 1.4 cores | 0.10 GB | 0.05 GB | 1.6 TOPS |
| Human Activity Recognition (HAR) is added to the body detector consumption. | 0.8 cores | 0.02 GB | 0.02 GB | 4.5 TOPS |
| UAD new objects (700×700) (10 FPS) | 3.5 cores | 0.28 GB | 0.09 GB | 5.6 TFLOPs |
Edge (Sophgo) test machine specifications
- OS: Ubuntu 20.04
- CPU: ARM Cortex-A53 @ 2.3 GHz: 8 cores
- RAM: 8 GB LPDDR4x
- VRAM: 8 GB LPDDR4x
- Disk: 64 GB eMMC
- NPU: 16 TFLOPS (FP16)
Sophgo support is available for a limited set of video analytics scenarios (without support for face recognition and phone detection). Acceleration on the integrated hardware accelerator is enabled by default; CUDA and TensorRT are not supported.
Calculate your server needs with the server capacity calculator.
Software
Supported OS: Windows 10+ (x86_64), Ubuntu 14+ (x86_64).
For stable operation of OMNI Agent, hardware resources should not be loaded above 80%.
If the database exceeds 50K persons, the agent’s working directory must be located on an SSD drive.
GPU requirements
System requirements
Linux / Windows:
GPU Driver >= 441.22
CUDA Toolkit 11.8
cuDNN v8.8.0 for CUDA 11.x
TensorRT 8.6 (optional)
Hardware requirements
- GPU with CUDA support (from NVIDIA GTX 1050 Ti up to RTX 4090 inclusive, Compute Capability 3.5–9.0).