Nvidia’s Jetson TX2 is perhaps one of the most powerful development boards available right now. The board continues the theme of its predecessor, the TX1, by packing extreme processing capabilities on the board itself. This means that developers working on Artificial Intelligence applications don’t need to depend on cloud processing for their calculations. The Jetson series from Nvidia has, since its inception, focussed on allowing creators to develop their projects without having to worry about accessing the right tools. The TX2 beefs up that claim by providing much better specs, a new architecture and a new development kit.
What is “edge” Machine Learning?
Edge Machine learning refers to the ability of an embedded system to perform complex and data intensive calculations locally. It is an alternative to cloud computing. The Nvidia Jetson TX2 removes the need for transferring data to a cloud and can be optimized for local, real-time processing.
Drones that need video processing in an instant, bots and self-driving cars that need faster computing to make quick decisions, sensitive data that needs to be protected and stored and processed locally can utilize this facility of the Jetson TX2.
What are the features of Nvidia’s Jetson TX2 and areas where it has improved over the TX1?
|Jetson TX2||Jetson TX1|
|GPU||NVIDIA Pascal, 256 CUDA cores||NVIDIA Maxwell, 256 Cuda cores|
|CPU||HMP Dual Denver 2/2 MB L2 + Quad ARM A57/2 MB L2||Quad ARM A57/2 MB L2|
|Video||4k x 2K 60 Hz Encode (HEVC)|
4k x 2K 60 Hz Decode (12-Bit support)
|4k x 2K 30 Hz Encode (HEVC)|
4k x 2K 60 Hz Decode (10-Bit support)
|Memory||8 GB 128 bit LPDDR4|
|4 GB 64 bit LPDDR4|
|Display||2x DSI, 2x DP 1.2 / HDMI 2.0 / eDP 1.4||2x DSI, 1x eDP 1.4 / DP 1.2 / HDMI|
|CSI||Upto 6 Cameras (2 Lane)|
CSI2 D-PHY 1.2 (2.5 Gbps/Lane)
|Upto 6 Cameras (2 Lane)|
CSI2 D-PHY 1.1 (1.5 Gbps/Lane)
|PCIE||Gen 2 | 1×4 + 1×1 OR 2×1 + 1×2||Gen 2 | 1×4 + 1×1|
|Data Storage||32 GB eMMC, SDIO, SATA||16 GB eMMC, SDIO, SATA|
|Other||CAN, UART, SPI, I2C, I2S, GPIOs||UART, SPI, I2C, I2S, GPIOs|
|USB||USB 3.0 + USB 2.0|
|Connectivity||1 Gigabit Ethernet, 802.11ac WLAN, Bluetooth|
|Mechanical||50mm x 87mm (400-pin Compatible Board-to-Board Connector)|
The biggest change from the TX1 is that the TX2’s GPU is based on Nvidia’s Pascal architecture. This architecture uses a new manufacturing process that makes it possible to fit in more transistors. The new Pascal architecture involves the 14nm FinFET node process. At the risk of over-simplifying the complex structure, fin field-effect transistors (FinFETs), due to their structure, are essentially better at current control. This property allows manufacturers to pack more processing capabilities in a smaller form factor while simultaneously reducing power consumption.
CPU and RAM
The CPU comprises a quad core A57 with a dual-core Denver 2.0 for high performance single threaded functions. The TX1 had an A53 along with the A57, but the A53 was unusable.
The RAM has been doubled by using four 32-bit channels as opposed to the four 16-bit channels of the TX1. Thus the 8GB low-power DDR4 memory also brings with it a higher bandwidth making the TX2 one of the fastest development boards out there.
Display and Camera interface
There are quite a few display options that users can choose. The two DSI ports for high-speed serial connection between a display module and the processor have been carried forward from the TX1. The HDMI 2.0 port and Embedded DisplayPort (eDP) have also been continued. The only difference is the addition of a v1.2 DisplayPort for external display.
The Camera Serial Interface has 12 lanes that allow a maximum of six cameras to be connected. It also uses a new version of the physical layer for data transfer, the MIDI D-PHY v1.2, an industrial standard to connect the cameras to the SoC. This version of the D-PHY bumps up the transfer rate per lane to 2.5Gbps. However, note that this is not the latest version of the standard. The latest version has a transfer rate of 4.5Gbps. Comparatively, the TX1 uses v1.1 of the standard and thus has a slower transfer rate of 1.1Gbps per lane.
Onboard storage has been doubled with the TX2 now featuring a 32GB eMMC storage. The Embedded MultiMediaCard (eMMC) storage is a cheap yet robust storage solution. However, like an SD card, it is not as fast or equipped as a full-fledged SSD. But don’t worry, there is an SATA interface available to expand the storage if you’d like to.
Regarding physical dimensions, the form factor of the TX1 has been continued in the newer iteration. Note that the development board is larger than the actual TX2 board. The standalone TX2 module is roughly the size of a credit card (50mm x 87mm). But the development board looks like a traditional motherboard.
Let’s talk about performance
The TX2 has two working modes. The efficiency mode, as the name suggests, prioritizes power consumption over peak performance. This mode allows the TX2 to work at the maximum processing level of the TX1 while consuming just half the power (7.5W) of what the TX1 would have consumed. That is an incredible development!
The performance mode pushes the TX2 into raw, unmitigated power and of course, the power consumption takes a backseat. In this mode, the TX2 works at twice the processing level of the TX1.
Brian Benchoff at Hackaday ran benchmark tests to compare the Jetson TX2 with the Raspberry Pi 3. Unsurprisingly, the TX2 emerged as the victor. It was four times faster than the Raspberry Pi. I have noticed a trend that over the last few years, every new development board, like Asus’s Tinker Board, is pitted against the Raspberry Pi. Perhaps it is because everyone is so familiar with the Raspberry Pi that it has assumed the role of providing a point of relative benchmarking.
Moving on to another comparison, the benchmarks released by Nvidia also indicate that the Jetson TX2 is faster than Intel’s Xeon Server CPU. The comparison was made on the basis of the two boards’ performance using the GoogLeNet deep image recognition network. The TX2, using the Jetson SDK, and running under 15W was able to beat the Xeon CPU, which was consuming 200W. Nvidia attributes this powerful performance to the combined efficiency of the Pascal GPU and the optimized Jetson SDK.
What about the new Jetson SDK?
Nvidia’s Jetson TX2 comes with a new Jetson Development Pack. The Jetpack SDK 3.0. This sweet bundle has all the tools that are accessible right out of the box. Some valuable tools worth mentioning are TensorRT, and cuDNN for deep learning applications, NVIDIA CUDA and OpenGL for GPU assisted processing and OpenCV for computer imaging. The complete list of available tools is listed here. https://developer.nvidia.com/embedded/jetpack
When and how can I buy it?
It’s available on Amazon and other major reseller outlets. You can use the advertisement below to buy it.
We get an affiliates fee for every purchase, which we use to buy ourselves candy and write more articles like this one. The development kit will cost you $599. The TX2 module card will be available for $399, but it won’t be available until Q2. The TX1 developer kit is now available for a reduced price of $499.
Sources and further reading
- MIPI D-PHY – Physical layer standard for camera connections
- MIPI DSI – High-speed serial interface standard for displays
- Hands-on Nvidia Jetson TX2: Fast Processing for Embedded Devices – Hackaday.com
- NVIDIA Jetson TX2 Delivers Twice the Intelligence to the Edge – In-depth information of Nvidia’s Jetson TX2
- Jetson TX2 – elinux.org