YOLOv8 from Ultralytics is the current standard for real-time object detection. The RTX 5060 Ti 16GB at our hosting runs dozens of simultaneous camera streams.
Contents
Setup
- Ultralytics YOLOv8, PyTorch 2.5, TensorRT 10.2
- Input: 640×640 RGB
- Single batch, timed per image
PyTorch FP16 FPS
| Model | Params | FPS | ms/frame | VRAM |
|---|---|---|---|---|
| YOLOv8n | 3.2M | 720 | 1.4 | 0.8 GB |
| YOLOv8s | 11.2M | 520 | 1.9 | 1.1 GB |
| YOLOv8m | 25.9M | 320 | 3.1 | 1.8 GB |
| YOLOv8l | 43.7M | 210 | 4.8 | 2.6 GB |
| YOLOv8x | 68.2M | 145 | 6.9 | 3.5 GB |
TensorRT FP16
| Model | FPS | Uplift vs PyTorch |
|---|---|---|
| YOLOv8n | 1,450 | 2.0x |
| YOLOv8s | 980 | 1.9x |
| YOLOv8m | 580 | 1.8x |
| YOLOv8l | 370 | 1.8x |
| YOLOv8x | 240 | 1.7x |
TensorRT roughly doubles FPS with no quality loss.
INT8 Quantised (TensorRT)
- YOLOv8m: 1,100 FPS (mAP drops ~1.5%)
- YOLOv8l: 680 FPS (mAP drops ~1.5%)
INT8 usable for surveillance where the small accuracy loss is fine.
Multi-Stream Capacity
Assuming 25 FPS per camera feed:
| Config | Max camera streams (25 FPS) |
|---|---|
| YOLOv8s TRT FP16 | ~39 streams |
| YOLOv8m TRT FP16 | ~23 streams |
| YOLOv8m TRT INT8 | ~44 streams |
| YOLOv8l TRT FP16 | ~14 streams |
For video analytics at scale (retail, traffic, industrial), this card is a strong value play – 30+ HD camera streams at YOLOv8m-level accuracy.
YOLOv8 on Blackwell 16GB
30+ HD camera streams, TensorRT-accelerated. UK dedicated hosting.
Order the RTX 5060 Ti 16GBSee also: YOLO guide, computer vision, PaddleOCR benchmark.
