Robotics kit runs AI code on Myriad X, Coral Edge TPU, and Elkhart Lake GPU

Robotics kit runs AI code on Myriad X, Coral Edge TPU, and Elkhart Lake GPU

Mar 8, 2021 — by Eric Brown

— 269 views

The “Fast Sense X Robotics AI Platform” runs Lubuntu and ROS on Intel’s Elkhart Lake with up to 2x Myriad X and 3x Coral Edge TPU M.2 accelerators for up to 20-TOPS AI.

Moscow-based startup Fast Sense Studios unveiled an 84 x 55 x 30mm, Intel Elkhart Lake based embedded computer comprised of 3x stacked boards designed for deployment on robots. The Fast Sense X Robotics AI Platform provides autonomous robots or drones with 3D navigation with depth estimation and obstacle avoidance, fusing data from multiple depth cameras and LIDAR sensors. The system also supports object and pose detection and web-based telepresence.

Fast Sense X Robotics AI Platform (left) and block diagram
(click images to enlarge)

The Fast Sense X Robotics AI Platform ships with pre-installed Lubuntu 18.04.5 LTS, a lightweight variant of Ubuntu Linux with a minimal LXDE/LXQT desktop. The stack also includes preconfigured Docker containers for the Intel Movidius Myriad X and Google Coral Edge TPU edge AI accelerators plus a ROS (Robot Operating System) node. The software includes inference code examples designed to work with SSD MobileNet.

The board-set is powered by an unnamed COM Express module featuring Intel’s 10nm fabricated Elkhart Lake platform. Elkhart Lake based COM Express modules we have seen include Avnet’s MSC C6C-EL, Congatec’s Conga-TCA7, Eurotech’s CPU-161-19 and CPU-161-20, and TQ’s TQMxE40C1 and TQMxE40C2.

The system incorporates the high-end, quad-core Atom x6425E clocked at 1.8GHz with 3.0GHz Turbo Burst. The Atom x6425E has a 2MB L2 cache and 12W TDP plus 500MHz/750MHz Intel UHD Graphics with up to 32EU Intel Gen11 technology.

This is the first product we have seen that taps the new deep learning inference and computer vision support provided on the Gen11 GPU. The system integrates up to 2x Myriad X and 3x lower-powered Edge TPU accelerators, which are deployed here via M.2 modules. Along with the GPU, these accelerators add up to a total of 20-TOPS AI performance running on less than 24W.

Fast Sense X Robotics AI Platform side views
(click images to enlarge)

Each accelerator can be driven by different algorithms running in different Docker containers enabling developers to mix and match for a custom AI fusion solution. Fast Sense poses the example of an autonomous robot with 3x cameras pointed in different direction, with each video stream processed with different neural nets for object detection, depth estimation, and semantic segmentation.

In this example, object detection can be handled by a single Coral Edge TPU module, which could process all three 30fps streams in less than 30 milliseconds (ms) with 8 ms SSD MobileNet inference time. The more demanding task of depth estimation might require dual Myriad X VPUs along with the Intel GPU. With 3x video streams, “you would get around 100 ms interference time which is 10fps,” says Fast Sense.

The task of segmentation would fall to the two remaining Edge TPUs. These would share devices between frames from different streams to achieve around 6fps on all three streams. Each Edge TPU can run MobileNetV2 at up to 400fps.

The Fast Sense X Robotics AI Platform is equipped with 16GB LPDDR4x at up to 4.267 MT/s plus up to 64GB UFS 2.0 flash. The 2x USB 3.1 and 6x USB 2.0 ports appear to be deployed via Type-C connectors. Other features include a DisplayPort as well as terminal plug connectors for 6x UART, I2C, and 1-Wire for controlling actuators.

The system offers a 7-35V input and up to industrial temperature support depending on which AI modules you use. Fast Sense is planning to support Intel’s upcoming Keem Bay follow-on to Myriad X and may support other AI accelerators. Customization services are also available.

Further information

No pricing or availability information was provided for the Fast Sense X Robotics AI Platform, although the CNXSoft story that alerted us to the product speculates a roughly $1,000 price. More information may be found on Fast Sense Studios’ website.