ximea
Already today the Imaging applications benefit from the latest NVIDIA mobile GPUs:
Jetson TX2 and AGX Xavier.
Nevertheless, general benchmarks can't answer the question about performance comparison between the two latest NVIDIA Jetson modules.
Still, this remains a very practical question for many various applications like aerial mapping, UAV, robotics, self-driving (autonomous) cars, etc.
To provide real numbers, we've done comparative studies with Fastvideo and their SDK, which has lots of image processing modules for camera applications.
The Fastvideo team has done time measurements for most frequently used image processing algorithms like demosaic, resize, denoise, jpeg encoder and decoder, jpeg 2000 codec, etc.
This is just a small part of Fastvideo SDK modules, though they could be valuable to understand the performance speedup with AGX Xavier.
The same images and parameters were then used for comparison.
Here you can clearly see how substantial the Xavier boost is - in many cases it allows you to switch from offline to the realtime mode of operation. This is also true for multiple camera systems connected to Jetson Xavier.
Jetson TX2 and AGX Xavier benchmarks
From this, it can be concluded that the performance increase is in the range of 1.7 - 3 for imaging applications on Jetson.
Quite often the results of RAW image processing are further used as the input for AI applications, which have also been significantly boosted by new Volta hardware cores on Jetson AGX Xavier.
Following embedded vision camera lines are currently supported:
xiQ, xiMU, xiC, xiX and xiSpec.
Here you can check the API for: Jetson TX1, TX2, AGX Xavier
Note: XIMEA demosaic image processing code isn't fully optimized for ARM so performance on color cameras with XI_RGB output is lowered. To get higher framerate - the application should select RAW image formats (RAW8, RAW16) which do not require image processing (especially demosaic).
Credentials
Fastvideo Blog:
https://www.fastcompression.com/blog/xavier-vs-tx2.htm