78 frames per second. Three simultaneous 25 FPS camera feeds with FPS to spare. The RTX 3090 turns YOLOv8 into a proper multi-camera surveillance or analytics server — and with 22 GB of VRAM left over, you can stack additional AI models on top. We ran the numbers on GigaGPU.
Raw Detection Performance
| Metric | Value |
|---|---|
| FPS (640×640) | 78 FPS |
| Latency per frame | 12.8 ms |
| Precision | FP16 |
| Input resolution | 640×640 (COCO) |
| Performance rating | Very Good |
Benchmark conditions: FP16 inference, batch size 1, YOLOv8m model on COCO-format input at 640×640.
22 GB of Unused VRAM
| Component | VRAM |
|---|---|
| Model weights (FP16) | 1.8 GB |
| Processing buffer | ~0.5 GB |
| Total RTX 3090 VRAM | 24 GB |
| Free headroom | ~22.2 GB |
The 3090’s 24 GB is almost entirely untouched by YOLOv8. That opens up genuinely powerful combinations: run detection alongside PaddleOCR for reading text on detected objects, add an LLM for natural language alerts, or pair with Whisper for audio-visual security monitoring. The 3090 handles all of it without breaking a sweat.
Scale Economics
| Cost Metric | Value |
|---|---|
| Server cost | £0.75/hr (£149/mo) |
| Cost per 1M frames | £2.67 |
| Frames per £1 | 374,532 |
Under three pounds per million frames, processing nearly 6.7 million frames per day. For context, that is about 77 continuous hours of 25 FPS footage analysed daily — far more than most single-site deployments require. The per-frame cost is marginally higher than the RTX 4060, but the 3090’s raw throughput makes it the clear choice for multi-stream setups. Full comparison on the benchmark page.
The Multi-Camera Server
At 78 FPS, the 3090 comfortably handles three 25 FPS camera feeds simultaneously. With frame subsampling to 15 FPS per feed (still smooth enough for most detection tasks), you stretch to five cameras on a single card. For warehouse monitoring, retail analytics, or traffic management systems, the 3090 is the practical choice: fast enough for real-time, flexible enough for complex pipelines, and priced at £149/mo. Detailed planning guidance: best GPU for YOLO.
Quick deploy:
docker run --gpus all -p 8080:8080 ultralytics/ultralytics:latest yolo detect predictBrowse: YOLOv8 hosting guide, all benchmarks, Flux.1 hosting, SD hosting.
Deploy YOLOv8 on RTX 3090
Order this exact configuration. UK datacenter, full root access.
Order RTX 3090 Server
