It’s an exciting time for small factor computing. As if the Raspberry Pi wasn’t enough of an all purpose machine, more powerful boards capable of incredible feats keep appearing.
The Jetson Nano from Nvidia is a recent addition to the ranks of super powerful machine learning enabled boards. What makes it special? Should you buy one? What is the Nvidia Jetson Nano all about?
What Is the Nvidia Jetson Nano?
The Jetson Nano is a Single Board Computer (SBC) around the size of a Raspberry Pi, and aimed at AI and machine learning. Seemingly a direct competitor to the Google Coral Dev board, it is the third in the Jetson family alongside the already available TX2 and AGX Xavier development boards.
Nvidia are leveraging their prowess for graphics processing power for these small computers, using parallel neural networks to process multiple videos and sensors simultaneously.
While all three Jetson boards aim to be accessible to all, the Nano is for both hobby and professional developers. The dev kit comprises two parts—a baseboard for connectivity, and a System On Module (SOM) for the actual processing units.
What Is System on Module?
System on Module refers to any development board which has all system-critical parts in a removable module. The Nano features a 260-pin edge connector to attach it to a baseboard for development.
Once development is over, the SOM can be removed and added into an embedded system with custom inputs, and a new SOM connects to the baseboard for further development.
If all of this sounds a little familiar, it is!
This is the same setup as the Google Coral Dev board, which is a similar size, and also aimed at embedded machine learning for hobbyists and professionals alike!
What Are the Jetson Nano’s Specs?
Nvidia has packed a lot into the Jetson Nano:
- CPU: Quad-core ARM® Cortex-A57 MPCore processor
- GPU: Nvidia Maxwell™ architecture with 128 Nvidia CUDA cores
- RAM: 4 GB 64-bit LPDDR4
- Storage: 16 GB eMMC 5.1 Flash
- Video: 4k @ 30fps encoding, 4k @ 60fps decoding
- Camera: 12 lanes (3×4 or 4×2) MIPI CSI-2 DPHY 1.1 (1.5 Gbps)
- Connectivity: Gigabit Ethernet
- Display: HDMI 2.0 or DP1.2 | eDP 1.4 | DSI (1 x2) 2 simultaneous
- PCIE/USB: 1 x1/2/4 PCIE, 1x USB 3.0, 3x USB 2.0
- I/O: 1x SDIO / 2x SPI / 6x I2C / 2x I2S / GPIOs
- Dimensions: 69.6 mm x 45 mm
- USB: 4x USB 3.0, USB 2.0 Micro-B
- Camera: 1x MIPI CSI-2 DPHY lanes (Raspberry Pi camera compatible)
- LAN: Gigabit Ethernet, M.2 Key E
- Storage: microSD slot
- Display: HDMI 2.0 and eDP 1.4
- Other I/O: GPIO, I2C, I2S, SPI, UART
What Can It Do?
It will come as a shock to nobody that Nvidia has produced a board suited well to visual tasks. Object recognition is a key focus here, and the Visionworks SDK has many potential applications in this field.
Rather than use a separate processing unit for machine learning tasks, the Jetson Nano uses a Maxwell GPU with 128 CUDA cores for the heavy lifting.
The Jetson Inference project features demos of a pre-trained neural network performing high-performance multiple object recognition in a variety of environments. Feature tracking, image stabilization, motion prediction, and multi-source simultaneous feed processing are all featured in the available demo packages.
Perhaps most impressive is the DeepStream technology featured in the above video. Running live analytics on eight simultaneous 1080p streams at 30fps on a small single board computer is incredible, and shows the potential power of the Nano’s hardware.
What Will It Be Used For?
Given its prowess for video analysis and small form factor, the Jetson Nano will almost certainly shine in robotics and autonomous vehicles. Many of the demos show these applications in action.
Given its power and size, it’ll also likely work in embedded systems which rely on facial and object recognition.
For Hobbyists like us? It seems to be a perfect blend of powerful machine learning possibilities in a factor familiar to anyone who has fiddled with a Raspberry Pi. While you can use machine learning frameworks like TensorFlow on a Raspberry Pi, the Jetson Nano is much more suited to the task.
What Else Can the Jetson Nano Do?
The Jetson Nano runs Ubuntu, though a specialized OS image is available from Nvidia featuring software specific to the platform. While the primary focus of the board is machine learning, this is Nvidia so you’d expect some graphical wizardry to be going on too.
You won’t be disappointed. Demos showing particle systems, real-time fractal rendering, and an array of visual effects would only until recently have been found on flagship desktop graphics cards.
Given that its video encoding is rated for 4k @ 30fps, and decoding at 60fps, it is safe to assume the Nano will be perfect for video applications too.
Jetson Nano vs. Coral Dev Board: Which Is Best?
It is tough to say which is the better board between the Google Coral Dev board and the Jetson Nano at this stage.
Google’s TensorFlow neural network is a dominant force in the field of machine learning. It would follow that Google’s own Edge TPU coprocessor might work better for applications of TensorFlow Lite.
On the other hand, Nvidia has already shown an impressive array of machine learning based demos for the Jetson Nano. This, alongside the impressive graphics the Nano is capable of make it a real competitor.
How Much Does Jetson Nano Cost?
Price is another aspect we haven’t covered yet. The Google Coral Dev board retails at $149.99 while the Jetson Nano is only $99. Unless the Coral Dev board can bring something unique to the table, hobbyists and small developers might find the extra $50 a hard stretch to justify.
There is currently no price for the SOM alone for either board, but I would imagine for most hobby developers this won’t be quite as important. From a commercial point of view, the performance/price contrast is going to be what makes the critical difference between the Jetson Nano and the Coral Dev board.
The Jetson Nano is available from Nvidia directly along with third party sellers.
Buy: Jetson Nano direct from Nvidia
Could It Replace My Raspberry Pi?
While the Google Coral Dev board is powerful, it doesn’t stack up to the Raspberry Pi in some ways. The Raspberry Pi is a great hobby computer for DIY electronics. It can also double as a desktop computerin a pinch.
Sure, the Coral Dev board is powerful, but their own documents advise against attaching a mouse and keyboard. The Coral’s custom OS is for SSH connections primarily. It is, however, likely capable of sustaining any variation of Linux. This puts it right back up there as a direct Pi competitor
There is a problem though. If you want a board for learning machine learning, but one that can also perform other daily tasks, why would you buy the Coral Dev Board?
The Jetson Nano supports a display port, and as previously mentioned has impressive video examples straight out of the box. The custom Ubuntu desktop will be familiar to many and the cheaper price point will make it an attractive prospect for many, even those uninterested in machine learning.
AI for Everyone
At this stage, it is hard to say which will be the better board. It’s also unknown which will be more accessible to home developers. I look forward to spending time with both the Coral Dev and Jetson Nano boards to get a definitive answer!
It’s an exciting time to be tinkering with SBCs! If you are new to it and want a place to start, get a Raspberry Pi and follow our ultimate getting started guide!
Explore more about: Google TensorFlow, Jetson Nano, Machine Learning, Raspberry Pi.