Leopard Imaging unveiled a compact “EdgeTuring” edge AI system that runs Linux on a quad -A7 Socionext SC2000 ISP SoC along with a 26-TOPS Hailo-8 M.2 AI module, dual [email protected] Sony IMX477 sensors, and AWS Kinesis services.
Leopard Imaging announced a dual-camera EdgeTuring mini-PC for edge AI applications that combines a Socionext imaging processor with Hailo’s up to 26-TOPS, 3-TOPS per Watt Hailo-8 M.2 AI Acceleration Module. The EdgeTuring system is designed for industrial automation, smart devices, smart retail, and other edge IoT applications.
EdgeTuring (left) and Socionext SC2000
(click images to enlarge)
The EdgeTuring camera system “consumes less power, performs at a higher level, and ensures greater reliability for video analytics and privacy at the edge than alternative solutions,” claims Leopard. The system provides an accuracy ranging from 95 percent to 99 percent for deep learning-based computer vision applications such as object detection and image classification, claims the company.
We have previously seen a far more powerful Socionext SoC deployed with the Hailo-8 NPU on Foxconn’s BOXiedge v2 AI edge server, a model that features a SynQuacer SC2A11 SoC with 24x Cortex-A53 cores. The EdgeTuring has a far more modest SC2000 “Milbeaut” SoC with 4x 650MHz Cortex-A7 cores, but with an ISP with up to a 1.2Gpixel/sec processing rate.
The SC2000 camera SoC enables 360-degree, real-time spherical stitching with up to 4x cameras, plus image stabilization without mechanical gimbals and rolling shutter correction. Other features include HDR capture and up to full HD slow motion video capture at 240fps.
The Hailo-8 M.2 AI Acceleration Module is an M.2 M-key 2242 module equipped with the Hailo-8 NPU, billed as the world’s highest performance AI M.2 module. The module provides PCIe Gen3 x4 while an upcoming, mini-PCIe model will offer the same NPU, but with PCIe Gen3 x1. The module was announced last fall along with a Hailo-8 Evaluation Board.
The Hailo-8 NPU uses a “proprietary novel structure-driven” Dataflow architecture that differs from the Van Neumann architecture used on most neural processors. The architecture achieves low-power memory access by implementing a distributed memory fabric combined with purpose-made pipeline elements.
— ADVERTISEMENT —
Although Leopard Imaging normally shows up on LinuxGizmos in regard to their embedded CMOS sensors on products such as SparkFun’s Jetson Nano-based JetBot AI Kit, the company opted for dual Sony IMX477 sensors, which we have seen on products such as the 12.3-megapixel, HD-resolution Raspberry Pi High Quality Camera. The Sony IMX477 Diagonal 7.857 mm (Type 1/2.3) sensors offer 1.55 x 1.55 μm pixel size and 3840 x 2160-pixel resolution.
The color sensors output at [email protected] with EIS and [email protected] fps with HDR. Only one of the sensors work with the Hailo-8 chip for AI processing. The sensors are accompanied by FOV 43.4° horizontal lenses.
(click images to enlarge)
The 93.2 x 80.7 x 38.0mm EdgeTuring device is further equipped with Ethernet, micro-HDMI, USB 3.0 Type-C, micro-USB 2.0, and micro-USB based UART ports. There is also a microSD slot and a 12VDC input. The 242-gram device has a 0 to 60℃ operating range.
A Linux OS works with two AWS Amazon Kinesis services: Amazon Kinesis Video Streams (KVS) and Amazon Kinesis Data Streams (KDS). We have seen KDS on Amazon’s Intel Cherry Trail based DeepLens machine learning camera where it streams video to AWS and Amazon Rekognition for video analytics. Here, the Kinesis services “create a seamless experience for customers to stream and analyze videos using a simple internet connection,” says Leopard.