Edge Computing: Chip Delivers High Performance Artificial Intelligence

Edge Computing: Chip Delivers High Performance Artificial Intelligence
Edge Computing: Chip Delivers High Performance Artificial Intelligence

The Hailo-8TM inference chip expands the number of industrial Artificial Intelligence (AI) applications possible for a wide range of industrial applications including optimization of production, processes, track & trace, logistics, quality, machine functions, and predictive maintenance by eliminating inherent limitations of server and cloud solutions with processing at the edge. Server and cloud AI solutions are suitable for a wide range of applications, but compute costs, network communication speed and latency factors pose limitations for many real time industrial and process applications. The Hailo-8TM inference chip and plugin modules expand the number of feasible applications.

Hailo was established in Israel in 2017 by members of the Israel Defense Forces’ elite technology unit, Hailo developed AI processors for edge devices. The company has received funding of $88 million including strategic investments from NEC & ABB. The team believes the Hailo technology is positioned to transform the AI chip industry, putting new compute possibilities within reach of more applications. The AI processor is designed to fit into smart devices in various industries and use cases including automotive, smart cities, retail and industry 4.0.

The Hailo-8 is another example of how real-time edge computing is increasing to deliver high performance and reliability in applications. Industrial automation started with central controls and, over time, processing has been pushed to the edge as technology advances made it practical. This has resulted in increases in availability, reliability, and performance, expanding application use cases. I had a discussion with Liran Bar, VP Business Development and Product Marketing at Hailo to learn more. 

The Hailo chip is a high-performance industrial grade Edge AI Processor

Liran Bar discussed other offerings in the market including Nvidia, Intel Myriad-X, and Google Edge TPU, noting that the Hailo-8 is the only chip optimized for low power and environmentally rugged edge applications including industrial and automotive. Bar provided detailed benchmark information supporting Hailo’s advantages in rugged edge applications.

“The Hailo-8TM processor is transforming visual intelligence and sensory perception for multiple industries by enabling smart devices to run neural network (NNs)-based applications more effectively at the edge," said Liran Bar, VP Business Developemtn and Product Marketing at Hailo. "Our M.2 and Mini PCIe high-performance AI acceleration modules expand these abilities by enabling customers to integrate AI capabilities into edge devices, providing a more flexible and optimized solution for accelerating a large range of Deep Learning-based applications with high efficiency, and optimizing time to market with a standard form factor.”

Add-on modules

In addition, the Hailo-8 can be applied using plugin add-on board modules delivered using the popular M.2 and mPCIe connector standards found in many computers including embedded industrial PCs adding high performance AI processing without degrading other applications in the computer. Specifications include 26 tera-operations per second (TOPS) Hailo-8™ processor with high power efficiency of 3 TOPS/Watt supporting industrial environment requirements.

This is analogous to early PCs coprocessor add-ons to achieve high performance floating-point mathematical calculation performance and video display coprocessors to achieve high resolution/performance graphics. For example, the original IBM PC included a socket for the Intel 8087 floating-point coprocessor (aka FPU), which was a popular option for people using the PC for computer-aided design or mathematics-intensive calculations.

System approach

Hailo devices are supported with the Dataflow Compiler, a complete & scalable software toolchain which seamlessly integrates with existing deep learning development frameworks enabling easy integration in existing development ecosystems. Open standards are supported including TensorFlow and ONNX.


AI optimized architecture

Liran Bar explain the architecture which is very intriguing since the application essentially drives the dynamic configuration of chip resources to optimize performance. The Hailo-8 chip contains many sets of processing elements each with control, memory, and compute which are configured and allocated to execute various layers of the neural network graph based on the user application as illustrated in the animation below. This is in contrast to the general purpose computer model based on the classic von Neumann architecture also known the Princeton architecture with a single CPU, memory and control.

The Hailo-8 chip has many sets of control, memory and compute elements that are configured and allocated to various layers of the neural network graph based on the user application for high performance.

It is not essential for users to understand the details of this, but there are a number of resources available on the Hailo website to learn more details.

Application example

Foxconn has been deploying several in-house developed AI solutions on different production lines, leading to an improvement in reporting accuracy from 95% to 99% and a reduction of at least one third of the operating costs for defect inspection projects.  Creating a standardized solution Foxconn combined its high-density, fan-less, edge computer, "BOXiedge™," with Socionext's high-efficiency parallel processor "SynQuacer™" SC2A11, and the Hailo-8™ deep learning processor to create an energy efficient standalone AI inference node for video analytics at the edge. The new product is capable of processing and analyzing over 20 streaming camera input feeds at the edge in real time for video analytics and privacy, including image classification, detection, pose estimation and various other AI-powered applications.

“Efficient deep learning processing at the edge will usher in a new generation of Industry 4.0 by ensuring manufacturers never have to compromise on performance, cost, or latency," Bar said. "Current solutions are often run on outdated processing architecture and cannot keep up with computing demands such as multi-device automation and connectivity that smart factories require. The Hailo-8 edge AI processor enables multiple streaming camera input feeds in real time, ensuring top performance for video analytics while maintaining privacy. We look forward to helping lead industrial 4.0 into a new digital age.”

Edge computing trend

The Hailo-8 is another example of how real-time edge computing is increasing to deliver high performance and reliability in applications. Industrial automation started with central controls and overtime processing has continually been pushed to the edge as technology advances made it practical resulting in increased availability, reliability, and performance expanding application use cases.



TensorFlow Tutorial.
But what is a Neural Network? | Deep learning, chapter 1

About The Author

Lydon brings more than 10 years of writing and editing expertise to Automation.com, plus more than 25 years of experience designing and applying technology in the automation and controls industry. Lydon started his career as a designer of computer-based machine tool controls; in other positions, he applied programmable logic controllers (PLCs) and process control technology. In addition to working at various large companies (e.g., Sundstrand, Johnson Controls, and Wago), Lydon served a two-year stint as part of a five-person task group, where he designed controls, automation systems, and software for chiller and boiler plant optimization. He was also a product manager for a multimillion-dollar controls and automation product line and president of an industrial control software company.

Read more

Did you enjoy this great article?

Check out our free e-newsletters to read more great articles..