Powered by a 128-core Maxwell GPU capable of 472GFlops at half precision, a 4-core ARM A57 CPU, with 4GB of LPDDR4 memory and 16GB of flash storage, the $130 Jetson Nano has been labelled as a low-powered AI computer.
For H.264 and H.265 video, the Nano is capable of processing eight 1080p streams in parallel while running object detection on all eight streams simultaneously at a rate of 30 frames per second.
Nvidia senior manager of product for autonomous machines Jesse Clayton said Edge TPU was "fast for a few small classification networks", but that the architecture was not suited for "large, deep neural networks".
"Jetson Nano, like all other Nvidia products, is architected around a general purpose GPU, which means the architecture is flexible enough to support the networks that researchers are still developing today," Clayton said.
"Jetson Nano is for full AI systems, it supports many, many hundreds of different networks, many popular frameworks, where the alternatives just don't."
The GPU giant has released a set of metrics that show the Edge TPU leaving Jetson Nano in its dust, but only on a pair of workloads. Nvidia is saying that the same software stack is used from Nano edge devices, all the way up to its larger, and almost 10 times as expensive, Jetson Xavier boards.
At the same time, Nvidia is also releasing a $100 Nano developer kit, which is aimed at developers, students, and the maker set. The dev kit version does not have onboard storage and requires an SD card.
The developer version will be available from Tuesday, with the full Nano available in the second quarter.
In December, Nvidia launched the Jetson AGX Xavier module, which has six processing units, including a 512-core Nvidia Volta Tensor Core GPU, an eight-core Carmel Arm64 CPU, a dual NVDLA deep-learning accelerator, and image, vision, and video processors.