NVIDIA’s Jetson TX2 is more than a worthy successor to the original. It’s a new way to do things.
Artificial Intelligence and machines that can learn are how the things we use every day will be improved. Google and Android are all-in with AI through Google Assistant and machine learning, so it’s important to know how the back end operates, how they got there and what types of equipment makes it all possible. And it’s really cool, too!
The people who will build this technology of the future will need the tools to do so. In 2017, NVIDIA is doing its part, and the Jetson TX2 is the embodiment of this idea. Developers need hardware that’s not only capable of doing the computing and thinking (yes, I’ll say it) that our smarter future is going to need, but is also easy to use and deploy.
AI at the Edge.
NVIDIA refers to this as “delivering AI at the Edge” and it’s an apt description. The TX2 is a complete supercomputer. It’s able to process data on its own at the place and time it’s actually happening instead of thousands of miles away via the internet. We take connectivity for granted because of the way we use it right now, but there are plenty of cases where waiting for a data round trip from a smart piece of machinery is just too long to wait. And a large part of this blue marble we live on doesn’t have a connection to the internet, and won’t for a very long time.
A small computer that can do just about anything and process all the data it collects itself is how you tackle these problems. NVIDIA seems to have nailed it here.
What is this thing?
This isn’t something you can find at Best Buy to use for things you do with your phone. It doesn’t run Android (but it certainly wouldn’t be difficult to fix that) and it’s something most of us won’t be buying. But it’s still a very important part of the things we love.
The Jetson TX2 is a development tool. The Jetson TX2 is also a field-ready module to power any AI-based equipment. It’s a computer the size of a credit card with all the inputs and outputs a “regular” computer has. When you plug the TX2 module into its specially designed backboard (that’s part of the development kit) it mostly turns into a typical small form factor PC complete with all the ports and plugs your desktop also has.
Developers can use this to build equipment around and actually use the Jetson itself to run demos and simulations. It’s a capable little machine that can do all the calculations something much bigger can do while using a minuscule amount of power to do so. The tech specs are impressive.
- NVIDIA Parker series Tegra X2: 256-core Pascal GPU and two 64-bit Denver CPU cores paired with four Cortex-A57 CPUs in an HMP configuration
- 8GB of 128-bit LPDDR4 RAM
- 32GB eMMC 5.1 onboard storage
- 802.11b/g/n/ac 2×2 MIMO Wi-Fi
- Bluetooth 4.1
- USB 3.0 and USB 2.0
- Gigabit Ethernet
- SD card slot for external storage
- SATA 2.0
- Complete multi-channel PMIC
- 400 pin high-speed and low-speed industry standard I/O connector
The best tech spec is that the Jetson TX2 is a pin for pin drop in replacement for last year’s Jetson TX1. Let that sink in for a bit — developers who are using existing NVIDIA TX1 computers to power the brains behind their equipment will be able to shut things down, pull the old board and put in the new one. The software for the TX1 will be updated to the same software the TX2 is using so it will literally be a drop in replacement. If you’ve ever done any type of field or factory work on equipment that costs a lot of money when it has any downtime, you understand how important this is. While the next generation equipment is being developed, it’s using hardware that works 100% with the existing generation.
The secret here is through NVIDIA’s Pascal GPU cores. The same reason Pascal cores are used in very high-end video cards designed for VR and 4K 3D gaming is why they’re used for the Jetson TX2. GPU cores are a more efficient way to crunch numbers. They’re faster and use a lot less power.
The holy grail of computing is artificial intelligence (AI): building a machine so intelligent, it can learn on its own without explicit instruction. Deep learning is a critical ingredient to achieving modern AI. Deep learning allows the AI “brain” to perceive the world around it; the machine learns and ultimately makes decisions by itself. It is now widely recognized within academia and industry that GPUs are the state of the art in training deep neural networks (DNN), due to both speed and energy efficiency advantages compared to more traditional CPU-based platforms.
NVIDIA GPU computers already do some amazing things. They drive the deep learning used for self-driving cars, teaching robots human-like motor skills such as walking and grasping, analyzing video at high-speed to provide text captions and even play Go. And beating really good human opponents.
GPU cores can do the same work using less power as traditional CPU computing.
The real test of AI and the brains that can drive it is on the horizon. Autonomous robots and drones are being developed for jobs like industrial inspection, portable medical devices that can be taken in the field to help those in need are desperately needed and even smart security cameras that can analyze what they are seeing and take appropriate action are soon to be realities. These ideas need computing that can drive AI with deep learning algorithms and the ability to analyze neural network collected data on their own. They can’t be attached to a cable and will be used in places where even Verizon has no coverage.
Besides being powerful, a computer designed to be small and portable has to be power efficient. Testing shows (.pdf file) that NVIDIA GPU-based computing can be equivalent to an Intel core i7 6700K CPU and use 6 watts of power compared to 60. For equipment that’s not connected to the power grid, that’s important.
We ran some benchmarks using AlexNet and GoogLeNet — CV based object category classification and detection testing software and the results were fantastic. In Max-P (high-power) mode, the Jetson TX2 was able to analyze an average of 641 images per second using the AlexNet Network while using just 13 watts of power. The GoogLeNet testing averaged 278 images per second while using 14 watts of power. Max-Q (low power) tests scored an average of 481 images per second on AlexNet and 191 images per second on GoogLeNet while using just 7 watts of power. This is just about twice what last year’s Jetson TX1 could deliver, and it was pretty good at it, too.
When you can process information this fast and this accurate on-site, a connection to the cloud isn’t the limiting factor it used to be.
In the lab
The Jetson TX2 should be very capable in the field. It’s the first of the next generation machines that will learn by doing without a connection to the cloud and a substantial upgrade from existing equipment. But it also has features that developers will love.
The credit card sized compute module can plug into a complete carrier board available as part of the Jetson TX2 development kit. The carrier board uses the 400 I/O pins on the Jetson module to provide standard desktop connections. A software developer can use a standard USB keyboard and mouse, a standard monitor and the Jetson TX2 to create a complete development environment.
Running on an Ubuntu 16.04 based Linux4Tegra operating system, all the tools you might need to develop and debug deep learning AI applications are included as part of NVIDIA’s JetPack software. Developers can download the package from NVIDIA’s Developer Zone as well as follow tutorials and community knowledge to see what the Jetson can do then begin work on their own ideas. Included software in the JetPack is pre-configured to run optimized on the TX2 processing system:
- cuDNN – a GPU-accelerated library of primitives for deep neural networks.
- NVIDIA VisionWorks is a software development package for Computer Vision (CV) and image processing.
- CUDA Toolkit – a comprehensive development environment for C and C++ developers building GPU-accelerated applications.
- TensorRT – a high performance deep learning inference runtime for image classification, segmentation, and object detection neural networks.
- NVIDIA Nsight Eclipse – A full-featured and customized Eclipse IDE for developing, debugging and profiling CUDA-C applications.
- Tegra System Profiler and Tegra Graphics Debugger – tools to profile and sample applications using OpenGL.
- The necessary collateral and assets to develop and design hardware using the NVIDIA Jetson TX2.
Using the same platform to build and debug any application is a must for anything intricate and complicated. It’s one of the ways developers can simplify the process and anything that can help make things easier makes for happier developers. While the Jetson TX2 may not be designed as the sole development and build computer any group would be using, knowing that it is capable is a boon for installation and field work. Making small adjustments and changes can be done on the Edge the same way the processing is without sending data back to another computer bank to process and return.
Equipment can be designed using the available hardware assets and drawings to not only reduce complexity but to allow an easy interface using readily available peripherals and software. Armed with a laptop and a USB cable, an engineer or field tech has everything needed to rebuild from the ground up if necessary.
The NVIDIA Jetpack software means developers can focus on their work, not setting up a build environment.
Even the installation of NVIDIA’s Jetpack is streamlined. Reviewers were provided with an updated version to install, and following a few simple instructions through a clever GUI had a complete rebuild of all the software finished with just a few steps and a cup of coffee. Again, we see NVIDIA making things easier so developers can focus on their work rather than maintaining the build environment itself.
You can actually build and debug software on the Jetson TX2, while having an assortment of other applications running to write a blog post.
After a few days of setting things up and testing everything, I came away very impressed with what NVIDIA is delivering here. The first Jetson TX1 was a great product that filled a need for fast development using GPU cores to do the heavy lifting for deep learning neural network applications. In a very short time, NVIDIA has raised the bar with a successor that can break the dependency on the cloud using the same familiar development tools and techniques.
The technology of the future will excite and inspire us all. Products like the Jetson TX2 are what will make that future possible. The NVIDIA Jetson TX2 Developer Kit is priced at $599 for retail orders and $299 for students.
See at NVIDIA Embedded Developers portal