So far, most applications of machine learning in the physical world have involved expensive equipment such as autonomous vehicles and industrial robots, so it hasn't mattered that the necessary computing power hasn't been particularly cheap.
But attention is turning to adding ML to mass-market products such as video surveillance systems, home robots, visual inspection systems for factories, and enterprise IoT projects.
Potential buyers are sensitive to price, putting earlier hardware "out of reach", according to Nvidia senior manager of product for autonomous machines, Jesse Clayton.
Now Nvidia has entered this part of the market with the Jetson Nano.
This 70x45mm module incorporates a 128-core Maxwell GPU plus a four-core ARM CPU, along with 4GB of RAM and16GB of eMMC storage, plus support for up to 12 camera inputs, two video inputs and an array of I/O interfaces.
The hardware is capable of decoding up to 16 streams of 720p video, eight 1080p streams, two 4K streams (all at 30fps) or one 60fps 4K stream. This means it can, for example, monitor eight full HD security cameras simultaneously, performing object detection on all of them.
Nvidia is highlighting two particular selling points in addition to the low (US$129) price.
Firstly, the Jetson Nano runs the same software as the rest of the Jetson family, including support for deep learning and machine vision, as well as higher-level functions such as depth estimation, object detection and gesture recognition.
Secondly, it runs a wider range of modern AI software than the Intel Neural Compute Stick 2 or the Coral Edge TPU, and in most cases it runs it more quickly. One exception is Google's MobileNetV2 computer vision software, which runs faster on the Edge at low resolution. However, the Jetson Nano can run it at higher resolutions (including 1920x1080).
The Jetson family delivers performance, power efficiency, ease of programming, and encapsulation (in the sense of being ready-to-use modules), said Clayton.
The Jetson Nano is an industrial quality product, but Clayton explained that Nvidia was well aware of the maker community's interest in ML.
So it has also released the US$99 Jetson Nano Developer Kit, available immediately from Nvidia's website and its distributors around the world.
It uses the same GPU and CPU, but incorporates USB, HDMI and Ethernet ports, and is compatible with various sensors and other devices designed for the Raspberry Pi, including those from Adafruit, Clayton said. Adafruit's Blinka library works with the kit.
No storage is included with the Jetson Nano Developer Kit; customers must add a suitable SD card.
Nvidia provides tutorials, project instructions, and access to a developers' forum and the Nvidia Deep Learning Institute.