Can an A100 GPU handle 8 parallel YOLOv11-L real-time inference streams at 1080p 30FPS?

Gaming and Visualization Technologies General Topics and Other SDKs GPU - Hardware
1

Hello,

I am working on a real-time object detection project in my company, and we plan to use YOLO for inference on multiple IP cameras. Currently, we do not have an on-premise GPU server, so we are considering purchasing a system equipped with an NVIDIA A100 GPU with this, I’m going to run the llm model in the future.

Our goal is to run 8 separate IP camera streams, each at:

  • Resolution: 1920 × 1080

  • Target inference speed: 30 FPS

  • Model: YOLOv11-L (Ultralytics)

Before making a purchase, I would like to confirm the following:

1. Can a single A100 GPU run 8 YOLOv11-L models in parallel at 1080p 30FPS?

If anyone has benchmark experience with YOLOv8/L11-L on A100, I would appreciate performance guidance—especially regarding GPU throughput, memory usage, and expected real-time stability across 8 streams.

2. If the A100 cannot reliably support this load, what edge system would you recommend?

We are open to alternatives such as:

  • Jetson AGX Orin industrial edge PCs

  • x86-based industrial computers with RTX A2000/A4000

  • Any proven hardware configuration capable of handling 8 channels of 1080p YOLO inference at 30FPS

If possible, we prefer a design suitable for factory environments.

Any recommendations, benchmarks, or architectural guidance would be extremely helpful.

Thank you in advance for your support.

2

What’s the input size of the model used during inference?

3

In my current implementation, I do not manually set the input tensor size.
The raw 1920×1080 frame is passed directly to the YOLO detector, and the model internally resizes it to its default inference size (typically 640×640 for Ultralytics YOLO).