Table of Contents
Why FPS Matters for Object Detection
Frames per second is the single metric that determines whether your YOLOv8 deployment can handle live video feeds. Running object detection on dedicated GPU hosting lets you control exactly how much throughput you get, but picking the wrong GPU means dropped frames and missed detections. For vision model hosting workloads like security camera analytics, autonomous vehicle pipelines, and manufacturing quality control, real-time performance is non-negotiable.
We benchmarked every YOLOv8 variant from nano to extra-large across five GPUs available on our UK-based servers. These are real inference numbers, not theoretical peaks.
Test Methodology
All GPUs tested on identical server configurations: AMD Ryzen 9 CPU, 64GB DDR5 RAM, 1TB NVMe, Ubuntu 22.04. Inference ran through Ultralytics YOLOv8 with PyTorch 2.1 and CUDA 12.1. Input: 640×640 images from the COCO validation set. Each benchmark processed 5,000 images and we report the median FPS across three runs. Batch size 1 unless otherwise noted.
We tested five model variants that span the full accuracy-speed tradeoff:
| Model | Parameters | FLOPs (B) | mAP (COCO) | FP16 VRAM |
|---|---|---|---|---|
| YOLOv8n (nano) | 3.2M | 8.7 | 37.3 | ~0.5 GB |
| YOLOv8s (small) | 11.2M | 28.6 | 44.9 | ~1.0 GB |
| YOLOv8m (medium) | 25.9M | 78.9 | 50.2 | ~2.1 GB |
| YOLOv8l (large) | 43.7M | 165.2 | 52.9 | ~3.4 GB |
| YOLOv8x (extra-large) | 68.2M | 257.8 | 53.9 | ~5.0 GB |
VRAM usage stays low across all variants, which means even budget GPUs can run the larger models. The bottleneck is compute, not memory. For a broader look at how these GPUs compare on other AI tasks, see our GPU comparisons category.
YOLOv8 FPS Results by GPU
All results measured at FP16 precision, batch size 1, 640×640 input resolution.
| GPU | VRAM | YOLOv8n | YOLOv8s | YOLOv8m | YOLOv8x |
|---|---|---|---|---|---|
| RTX 3050 | 6 GB | 195 FPS | 112 FPS | 58 FPS | 22 FPS |
| RTX 4060 | 8 GB | 310 FPS | 185 FPS | 96 FPS | 38 FPS |
| RTX 3090 | 24 GB | 420 FPS | 260 FPS | 140 FPS | 56 FPS |
| RTX 5080 | 16 GB | 580 FPS | 365 FPS | 198 FPS | 82 FPS |
| RTX 5090 | 32 GB | 740 FPS | 470 FPS | 258 FPS | 110 FPS |
The RTX 5090 delivers roughly 3.8x the throughput of the RTX 3050 on YOLOv8n, and the gap widens on heavier models. For production video analytics, the RTX 3090 remains the value leader at these price points. Compare it directly against newer cards in our RTX 3090 vs RTX 5090 analysis.
Batch Size Impact on Throughput
Single-frame inference is the standard for live video, but batch processing matters for offline workloads like analysing recorded footage. Higher batch sizes improve GPU utilisation and push total throughput significantly higher.
We tested YOLOv8m across batch sizes 1, 4, 8, 16, and 32 on each GPU:
| GPU | BS=1 | BS=4 | BS=8 | BS=16 | BS=32 |
|---|---|---|---|---|---|
| RTX 3050 | 58 FPS | 78 FPS | 84 FPS | OOM | OOM |
| RTX 4060 | 96 FPS | 138 FPS | 155 FPS | 162 FPS | OOM |
| RTX 3090 | 140 FPS | 215 FPS | 268 FPS | 310 FPS | 335 FPS |
| RTX 5080 | 198 FPS | 305 FPS | 375 FPS | 420 FPS | 448 FPS |
| RTX 5090 | 258 FPS | 402 FPS | 498 FPS | 570 FPS | 620 FPS |
OOM = out of memory. The RTX 3050’s 6GB VRAM limits batch processing above BS=8 on YOLOv8m. The RTX 3090 with 24GB handles BS=32 comfortably, delivering a 2.4x throughput boost over single-frame mode. If you need to choose between the 4060 and 3090, our RTX 4060 vs RTX 3090 comparison covers more workloads.
Real-Time Threshold Analysis
Real-time object detection typically requires a minimum of 30 FPS for smooth tracking. For applications like autonomous driving or drone navigation, 60 FPS is the target. Here is which GPU hits each threshold per model variant:
| Model | 30 FPS (minimum) | 60 FPS (smooth) | 120 FPS (high-speed) |
|---|---|---|---|
| YOLOv8n | All GPUs | All GPUs | RTX 4060+ |
| YOLOv8s | All GPUs | RTX 4060+ | RTX 3090+ |
| YOLOv8m | RTX 4060+ | RTX 3090+ | RTX 3090+ |
| YOLOv8x | RTX 4060+ | RTX 5080+ | None (single GPU) |
The RTX 3050 cannot sustain 30 FPS on YOLOv8x, making it unsuitable for heavy detection models in live pipelines. Every other GPU clears the 30 FPS bar on all variants. If your use case also involves OCR alongside detection, see our PaddleOCR hosting page for combined pipelines.
Cost per Million Frames
Raw FPS doesn’t account for hosting cost. Using our GPU server pricing, here is the cost to process one million frames with YOLOv8m at batch size 1:
| GPU | FPS (YOLOv8m) | Frames/day | Cost per 1M frames |
|---|---|---|---|
| RTX 3050 | 58 | ~5.0M | ~$0.60 |
| RTX 4060 | 96 | ~8.3M | ~$0.42 |
| RTX 3090 | 140 | ~12.1M | ~$0.25 |
| RTX 5080 | 198 | ~17.1M | ~$0.29 |
| RTX 5090 | 258 | ~22.3M | ~$0.31 |
The RTX 3090 again dominates on cost per frame, matching the pattern we see in cost per 1M tokens for LLM workloads. The older Ampere card’s lower hosting cost offsets its speed gap against Blackwell. For LLM workloads running alongside your detection pipeline, check our best GPU for LLM inference guide.
Deploy YOLOv8 on Dedicated GPUs
RTX 3090, RTX 5080, and RTX 5090 servers ready for real-time object detection. Full root access, NVMe storage, 1Gbps uplink from our UK datacenter.
Browse GPU ServersGPU Recommendations by Use Case
Single-camera live detection (security, retail):
- RTX 4060 — Clears 30 FPS on all models and handles up to 4 concurrent YOLOv8s streams at 640×640
- Budget-friendly entry point for vision model hosting
Multi-stream video analytics (warehouse, traffic):
- RTX 3090 — Best cost per frame, 24GB VRAM supports batch processing and concurrent model loading
- See more benchmarks on our benchmarks hub
High-speed detection (autonomous systems, drones):
- RTX 5080 or RTX 5090 — Only cards exceeding 60 FPS on YOLOv8x at full resolution
- Pair with Whisper for audio analysis on the same server; see our Whisper hosting page
All GPUs above are available on our dedicated GPU hosting platform with same-day deployment from our UK datacenter.
