NVIDIA Jetson TX2 Linux Benchmarks

Written by Michael Larabel in Processors on 14 March 2017. Page 2 of 4. 37 Comments

For today's launch testing article, I have results compared to just the Jetson TX1. For many of the tests, the Jetson TX2 was tested with its Max-P and Max-Q operating modes. In follow-up articles will be comparisons of the Jetson Tegra X2 performance to other ARM and x86 hardware, including the older Jetson TK1. Just for today's article, I was a bit short on testing time. First up are some Caffe and CUDA results with the TX1 and TX2.

With Caffe AlexNet using CUDA and cuDNN, the Jetson TX2 in its lower-power Max-Q mode was still 13% quicker than the Jetson TX1. With the Max-P performance mode, the TX2 was 32% quicker than the Jetson TX1.

Here's a look at the reported CPU and GPU power consumption as exposed via sysfs with the Jetson TX2 developer board. Unfortunately I don't have any comparison numbers to the Jetson TX1 since my original JTX1 board lacks power sensors, but NVIDIA is reportedly sending over an updated JTX1 that does expose power sensors.

And the exposed system power consumption of 6.4 Watts in the Max-Q mode and 8 Watts in the Max-P mode.

The SoC temperature while running AlexNet was 33~35°C. The Jetson TX2 fan only turns on when needed and spent most of the time off. The TX2 is running much cooler than the TX1.

The GoogleNet results also show a significant advancement in performance over the Jetson TX1.

The power and thermal numbers should be of more value once being able to compare it to the Jetson TX1, and the Phoronix Test Suite can then generate performance-per-Watt results. The perf-per-Watt will be interesting to compare to other ARM/x86 competition.

Related Articles