Raspberry Pi-style Jetson Nano is a powerful low-cost AI computer from Nvidia

Developers who want to use machine learning on homemade gadgets or prototype appliances just got a powerful new low-cost option, with Nvidia revealing the Jetson Nano.

The $99 Jetson Nano Developer Kit is a board tailored for running machine-learning models and using them to carry out tasks such as computer vision.

Nvidia has shown the board being used to highlight people and cars captured by CCTV streams, running real-time object detection on eight 1080p30 streams simultaneously, using a ResNet-based model running at full resolution and handling a throughput of 500 megapixels per second.

The tiny board packs an Arm-based CPU and Nvidia GPU, based on the 2014 Maxwell architecture, which together deliver 472 GFLOPs of compute performance and consume as a little as five watts.

SEE: Free machine learning courses from Google, Amazon, and Microsoft: What do they offer? (Tech Pro Research)

Nvidia released a series of benchmarks showing the Jetson Nano outperforming competitors when running various computer vision models. The results show the Jetson Nano beating the $35 Raspberry Pi 3 (no mention of the model), the Pi 3 with a $90 Intel Neural Compute Stick 2, and the newly released Google Coral board that uses the Edge TPU (Tensor Processing Unit). These tests involved running a range of computer vision models carrying out object detection, classification, pose estimation segmentation and image processing. Specifically, the Jetson showed superior performance when running inference on trained ResNet-18, ResNet-50, Inception V4, Tiny YOLO V3, OpenPose, VGG-19, Super Resolution, and Unet models.

The Jetson Nano was the only board to be able to run many of the machine-learning models and where the other boards could run the models, the Jetson Nano generally offered many times the performance of its rivals.

Nvidia’s senior manager of product for autonomous machines Jesse Clayton told TechRepublic’s sister site ZDNet that Jetson Nano’s GPU could run a broader range of machine-learning models than the specialist silicon found in Google’s Edge TPU.

However, it wasn’t a clean sweep for the Jetson Nano, with Google’s Coral board beating the Jetson Nano when running a trained SSD Mobilenet-V2 model handling 300×300 resolution images, with the Coral able to run at 48 frames per second (FPS), compared to 39FPS on the Jetson Nano.

The tests above are also Nvidia-supplied benchmarks, and in Google’s own testing of the Coral board it claimed the ability to “run MobileNet v2 at 100+ FPS, in a power efficient manner”. Online posts are also starting to emerge from Google Coral owners claiming it should outperform the Jetson Nano when running MobileNet v2 by a greater margin than Nvidia is claiming.

The Jetson Nano can also be used to train machine-learning models, giving it an advantage over Google’s Edge board, which also requires you to upload your model to Google for compilation. Training performance is likely to be limited, however, given the cost of the board, particularly compared to using more expensive PC GPUs or cloud-based GPU arrays, and Nvidia itself says that training should only be carried out by those who are “willing to wait longer for results”.

The Jetson Nano is built around a quad-core 64-bit Arm-based CPU, a 128-core integrated Nvidia GPU and 4GB LPDDR4 memory.

The new JetPack 4.2 SDK provides a complete desktop Linux environment for the board, based on Ubuntu 18.04, with the OS bundling the NVIDIA CUDA Toolkit 10.0, and libraries such as cuDNN 7.3 and TensorRT 5.

The SDK includes the ability to natively install popular open source ML frameworks, such as TensorFlow, PyTorch, Caffe, Keras, and MXNet, along with frameworks for computer vision and robotics development like OpenCV and ROS.

Nvidia says a range of peripherals can be hooked up to the Jetson Nano via its ports and GPIO header, such the 3D-printable deep learning JetBot that NVIDIA has open-sourced on GitHub, while the Raspberry Pi Camera Module v2 is also supported and can be connected to the board’s MIPI CSI-2 port.

For those getting started with machine learning on the Jetson Nano, Nvidia offers the Hello AI World guide, which it says will allow new users to have trained real-time image classification and object detection running on the board within a “couple of hours”.

Firms wanting to build the Jetson Nano into a finished product can buy it in a 70 x 45mm System on Module (SOM) form factor.

The 260-pin SODIMM-style SOM will start shipping in June 2019 for $129 — for an 1000-unit order. The production-targeted compute module will include 16GB eMMC onboard storage and enhanced I/O with PCIe Gen2 x4/x2/x1, MIPI DSI, additional GPIO, and 12 lanes of MIPI CSI-2 for connecting up to three x4 cameras or up to four cameras in x4/x2 configurations.

Jetson Nano Developer Kit specs