A whitepaper is available on Tegra X1 and Nvidia drive PX available here

http://international.download.nvidia.com/pdf/tegra/Tegra-X1-whitepaper-v1.0.pdf

The platform contains two tegra X1 processors delivering peak 2 TFLOPS

There’s also a video of the CES presentation available here

https://www.youtube.com/watch?v=zsVsUvx8ieo

From what I can see the HoG SVM is working at about 3Hz (12 detections counted in a few 5sec bursts fo around 2.4 detections/sec) on 1080p video

They say they have a 30Hz camera input at 6:50 and they appear to be running at least 3 detectors in parallel detecting multiple 60Hz signs and speed cameras on a motorway gantry

This part shows 4 detectors in parallel detecting total 8 vehicles in a 1080p frame

Training 40hrs of 1080p video on a Tesla K40 desktop GPU takes 16 hrs (01:50)

Touted applications include self-parking

http://blogs.nvidia.com/blog/2015/01/04/nvidia-drive-px-demo/

Given the 2 TFLOP capability of the dual X1 system the performance of a few Hz at 1080p is rather underwhelming pointing to very high FLOPS being required for the CNN implementation