Back

 Industry News Details

 
NVIDIA's Jetson TX2 Takes Machine Learning To The Edge Posted on : Mar 22 - 2017

Small development boards have become a key enabler of a recent wave of hardware startups. The most popular boards such as the Arduino and Raspberry Pi have inspired many projects and many knock-offs. These boards are great to kick start a project like a beer brewing machine, but they are limited in compute power. When your project needs real compute power, and possibly some local machine learning, where do you go? One popular choice is the NVIDIA Jetson boards. The first board was the TX1 released in November, 2015.  Now NVIDIA has released the more powerful and power-efficient Jetson TX2 board. The Jetson boards are siblings to NVIDIA’s Drive PX boards for autonomous driving and the TX2 shares the same Tegra “Parker” silicon as the Drive PX2. There are many synergies between the two families as both can be used to add local machine learning to transportation. The Drive PX boards are designed for automotive with extended temperature ranges and high reliability requirements. The Jetson boards are optimized for compact enclosures and battery power for smaller, portable equipment.

With devices such as robots, drones, 360 cameras, medical, etc., Jetson can be used for “edge” machine learning.  The ability to process data locally and with limited power is useful when connectivity bandwidth is limited or spotty (like in remote locations), latency is critical (real-time control), or where privacy and security is a concern.

For the Jetson TX 2 launch, NVIDIA hosted an event in San Francisco that brought in a number of companies and researchers using Jetson in their products or projects. The star of the show was the new Cisco Spark Board, which combines a large touch sensitive monitor that acts as an interactive white board can act as a video conferencing system that recognizes and adapts to the people in the room. The board is capable of tracking the number of people in a room and adjusting the display zoom and offset accordingly. It also implements a 12-element microphone array to track who is talking and highlight that person. In addition, the board can also automatically wake up when an authorized person enters the room and resume their workflow from where it was last. Combining white board, collaborative tools, and video conferencing into one screen makes the system more effective and more productive. The integrated Jetson supports the 4K resolution display of the Spark Board and can assist local facial recognition.

One of my favorite products on display using the Jetson board is a portable handheld 3D scanner from Artec 3D. The “Leo” handheld scanner can be used to create a 3D model of any object up to the size of a car. An older version was used to create a 3D model of President Obama. The Jetson board is used to calculate and display the part so the object that have successfully been models on a display on the back of the scanner. The images form a point-cloud of the object that is later turned into a set of geometry and textures to build a 3D model in the computer. The level of detail is set by the user. The end result is a detailed 3D model that can be used for art and design, including CGI, healthcare, industrial design, science and education, etc.

 Another interesting system using Jetson was an intelligent drone from Enroute for automated search and rescue operations and maintenance and inspection work. The drone uses deep learning on Jetson for object detection, path planning, and collision avoidance. MIT was represented with a Jetson-powered mobile research robot. Toyota also had a “human support” research robot that uses Jetson.

Key advantages of the Jetson TX2 over the original TX1 are that it adds two additional, higher performing Denver CPU cores to the four Cortex-A57 cores in the TX1, NVIDIA’s latest Pascal GPU, and it offers twice the memory capacity and bandwidth. (See table below for more detail.) Because the Jetson TX2 is based on the same Tegra chip used in the NVIDIA Drive PX2 platform, it uses TSMC’s automotive grade 16nm FinFET process. The newer process node helps lower power from the 20nm process used on the TX1. The TX1 had a 10W power rating, but the TX2 can offer similar performance at less than 7.5W (which NVIDIA calls Max-Q). The TX2 is also capable of running at full speed at 15W (Max-P) with twice the compute performance of the TX1. The CPU clock speeds are up to 2GHz and the Pascal GPU can run up to 1.3GHz.

The company gives developers a jump start with the JetPack software development kit (SDK) for machine learning and GPU compute that taps into the company’s extensive AI ecosystem. The SDK supports Google’s Tensorflow, as well as NVIDIA’s cuDNN and CUDA libraries. The development board OS is Ubuntu Linux. The latest JetPack 3 SDK includes a developer’s preview of Ubuntu with kernel 4.4.Source