NVIDIA Jetson AGX Xavier Benchmarks


Fig.1. Jetson AGX Xavier Developer Kit

With the help of NVIDIA Jetson AGX Xavier developer kit, many customers already are developing and deploying end-to-end AI robotics and edge applications in the fields of aerial inspection, city mapping and more.

The AGX Xavier is supported by NVIDIA JetPack and DeepStream SDKs, as well as CUDA®, cuDNN, and TensorRT software libraries.
The power of the new NVIDIA Xavier processor offers more than 20X the performance and 10X the energy efficiency of its predecessor, the NVIDIA Jetson TX2.

The dev kit provides the tools needed to get started right away and XIMEA team tested various cameras with the results listed below.

AGX Xavier in multiple camera setup with 3 cameras

Hardware components:

  • Jetson AGX Xavier Developer Kit
  • 2x MC031CG-SY XIMEA camera: 3.1 MP, 2064 x 1544, Color model with USB3 interface
  • MX124CG-SY XIMEA camera: 12.4 Mpix, 4112 x 3008, Color model with PCIe interface
  • GPU desktop as a receiving station

Standard full frame processing pipeline:

  • 8/10/12-bit acquisition
  • color pre-processing
  • H.264/H.265 encoding, RTSP streaming

Benchmarks with multi camera setup

First MC031 model Second MC031 model MX124 model
8-bit 60 Fps 60 Fps 60 Fps
12-bit 60 Fps 60 Fps 45 Fps
Processing latency 30 ms 30 ms 45 ms



Fig.2. NVIDIA AGX Xavier with Connect Tech Rogue carrier and XIMEA MX089CG-SY-X2GX-FF camera

AGX Xavier with high speed camera

Hardware components for second camera model setup:

  • NVIDIA AGX Xavier with Connect Tech Rogue carrier
    • FireFly to M.2 adapter to connect the camera
  • XIMEA MX089CG-SY-X2GX-FF camera: 8.9 MP, 4112 x 2176, Color model with PCIe interface and FireFly connector
  • GPU desktop as a receiving station

Benchmarks with MX089 camera

Full HD, 1920 x 1080, 2.1 MP images (ROI on the camera)

  • Working frame rate: 60 fps
  • Image processing time: 12.5 – 13.6 ms (pp. 3-10 from processing pipeline)
  • Memory utilization: 10.6% of 32 GB
  • GPU utilization: 10-20%
  • CPU utilization: ~10%
  • Encoding bitrate: 9800 Kbps
  • G2G latency: 50-70 ms

4K UHD, 3840 x 2160, 8.3 MP images (ROI on the camera)

  • Working frame rate: 60 fps
  • Image processing time: 25.5 – 27.5 ms (pp. 3-10 from processing pipeline)
  • Memory utilization: 12.5% of 32 GB
  • GPU utilization: 30-60%
  • CPU utilization: 15-20%
  • Encoding bitrate: 9800 Kbps
  • G2G latency: 70-90 ms

AGX Xavier Acquisition Module for MIPI cameras

  • NVIDIA Jetson AGX Xavier Developer Kit with NVIDIA Jetson AGX Xavier Module 32GB
  • 4x 5664x 4248, 24MP monochrome MIPI image sensors

Processing:

  • 4x sensor acquisition at 30 FPS
    • Transfer to AGX memory
  • Render on a screen of images from one image sensor at 30 FPS