Nvidia is in the machine Learning and AI play for a number of years. The company launched the Jetson TX1 Supercomputer on Module in 2015 as an embedded solution for robots, drunk and self-driving vehicles that need to do a lot of visual computing. It was the start of a whole series of "NI" AI "successful" products. Nvidia says there are hundreds of thousands of Jetson developers today. While this was a workable solution for commercial businesses, the $ 599 prize meant it was often too expensive for makers, hobbies, and amateur enthusiasts.
Today it's all changed with the launch of the Jetson Nano, a $ 99 AI development package that opens the way for a Raspberry Pi-like revolution – this time for machine learning.
The secret sauce in Nvidia's AI products is, of course, its GPUs. The Jetson TX1 uses a 1024 GFLOP Maxwell GPU with 256 CUDA cores. The TX2 offers 1.3 TFLOPs with a 256-core Pascal GPU, and the top-of-the-range Jetson AGX Xavier breaks 10 TFLOPs with its 512 core Nvidia Volta GPU . But the Jetson AGX Xavier also breaks the $ 1000 barrier! For the $ 99 Jetson Nano, Nvidia has opted for a 128 CUDA core GPU based on Maxwell architecture. It offers 472 GFLOPs.
GPU support is a 64-bit square arm Cortex-A57-based CPU, 4GB RAM, a video processor. It can decode up to 4K 30fps encoder or 4K 60fps – and support for PCIE and USB 3.0.
The video capabilities of the Jetson Nano are impressive. The idea is not that you can watch 4K video, but rather that the unit can process multiple video streams (think of drones with multiple cameras) for object detection, tracking and autopsy. While 4K 60fps look good, the Jetson Nano is able to decode 8 video / camera feeds at Full HD at 30 frames per second! Once decoded, the streams can be processed simultaneously by the object learning algorithms, etc.
The Jetson Nano comes in two forms. A Module – measuring just 70 x 45 mm – for use in final production ready designs, and a development package that looks like & # 39; n Raspberry Pi and offers a key solution for developers and enthusiasts. The former comes with 16GB eMMC on board, while the latter comes with microSD card.
Unlike previous Jetson platform iterations, Nvidia provides two separate (but related) uses of the Jetson Nano. On the one hand, the development package will be useful for commercial organizations that want to build products with machine-learning capabilities. The product can be designed using the development package and then the modules are used for the final product. So the other Jetson boards and modules are used. The second use case is for enthusiasts and hobbyists who never use the module version, but are happy to create development package-based projects, such as the Raspberry Pi.
That's why Nvidia is ready to sell both the modules and the development packages, not just through wholesale suppliers, but to a larger market through more conventional outlets.
The Raspberry Pi uses a four-core Cortex-A53-based processor and comes with a maximum of 1 GB of RAM. Although it can be fun to use simple Python scripts and other basic tasks, it can be painful to use a desktop environment. The Jetson Nano has a four-core Cortex-A57-based CPU and 4 GB of RAM. This should mean that it is at least twice as fast as the Raspberry Pi for non-machine learning tasks. Plus, the extra RAM should allow a desktop environment to run smoother.
In addition, the Jetson Nano comes with 40 GPIO pens, just like the Raspberry Pi. While Nvidia does not specifically specify Raspberry Pi compatibility, it says the Jetson Nano "is compatible from the box with many peripherals and other add-ons." There is also support for the Adafruit Blinka library and the Raspberry Pi camera V2. The board starts with a full Linux desktop environment via Linux4Tegra, which is derived from Ubuntu 18.04.
In other words, the Jetson Nano is just like a Raspberry Pi, but better, stronger, faster! Add all the ML goodness on top and you have a potential game changer.
To demonstrate the board's capabilities, Nvidia launches the JetBot, an open source AI project based on Jetson Nano. It comes down to a piece of material, hardware setup guide, and tutorials. The idea is that someone with some basic Python skills should be able to build the small robot and learn everything about motor control, camera image acquisition and AI training by learning JetBot to follow objects, avoid collisions, and so on.
One of the reasons that Raspberry Pi was so successful, compared to other poor-based Single Board Computers, is that the software is always up-to-date. There are too many boards that offer initial support for a version of Linux and then the distribution is never updated or upgraded. No security solutions, no new packages, and certainly no new core versions.
Nvidia understands this and does a good job of keeping it active and relevant. The Jetson TX1 supports Linux 3.10 and uses Ubuntu 14.04. Over time, support fort kernel 4.4 was added, followed by kernel 4.9. Similarly, the base Ubuntu has been upgraded from 14.04 to 16.04 and now 18.04 upgraded.
This means that Nvidia offers a unified development environment across all its Jetson boards. You can start developing a project on the Jetson Nano, but if you need more GPU power, & nbsp; an upgrade to a more advanced Jetson board will have little or no penalty & # 39; a software perspective.
It looks like the Jetson Nano can be a fantastic board. The price is good, overall computer performance is significantly better than Raspberry Pi. The machine learning features (both software and hardware) are excellent, and the potential compatibility with existing hats and sensors means that hobbyists can use (and improve) existing projects. I need to get my hands on a board very soon, so watch out for a full review here and there the Gary explains YouTube channel.