Artificial Intelligence (AI) has been making headlines over the past few months with the widespread use and speculation about generative AI, such as Chat GPT. However, AI is a broad topic covering many algorithmic approaches to mimic some of the capabilities of human beings. There is a lot of work going on to use various types of AI to assist humans in their various activities. Note that all AI has limitations. It generally doesn’t reason, like we do and it is generally best at recognizing patterns and using that information to create conclusions or recommendations. Thus, AI must be used with caution and not used to form conclusions beyond what it is capable of analyzing. Also, AI must be tested to make sure it is not biased based upon the training sets used to create the AI algorithms.
AI comes in big and small packages. Much of the training is done in data centers, where vast stored information and sophisticated processing of that information can be used to train AI models. However, once a model is created it can be used to interpret information collected locally via sensors, cameras, microphones and other data sources to make inferences that can help people to make informed decisions. One important application for such AI inference engines is for mobile and consumer devices that may be used for biometric monitoring, voice and image recognition and many other purposes. Inference engines can also be used to improve manufacturing and various office tasks. AI inference engines are also an important component in driving assistance and autonomous vehicles, enabling lane departure detection and collision avoidance, and other capabilities.
There have been some recent announcements for AI inference engine chips from Brainchip and Hailo. Let’s look at these announcements and their implications for processing and interpreting vast amounts of stored and sensed data. Slides from a Brainchip briefing provided some economics for making and potential applications for AI. It said that the costs of training a single high-end model is about $6M and annual losses in manufacturing due to unplanned downtime were about $50B. In terms of the opportunity, about 1TB is generated by an autonomous car per day and there is about a $1.1T cost of healthcare and lost productivity due to preventable chronic disease. A PWC report projects $15.7T in global GDP benefit from AI in 2030 and Forbes Business Insights projects $1.2T in AI internet of things (AIoT) revenue by that year.
To help enable this opportunity, Brainchip announced its 2nd generation of its akida IP Platform. The figure below from the briefing seems to show that this platform may be using chiplet technology that integrates the Akita Neuron Fabric chiplet to perform various functions.
The akida IP Platform is a digital neuromorphic event-based AI device that is capable of some learning for continuous improvement. The company says that its second-generation of Akida now includes Temporal Event Based Neural Nets (TENN)spatial-temporal convolutions that supercharge the processing of raw time-continuous streaming data, such as video analytics, target tracking, audio classification, analysis of MRI and CT scans for vital signs prediction, and time series analytics used in forecasting, and predictive maintenance. The second-generation device also supports Vision Transformers (ViT acceleration, a neural network that is capable of many computer vision tasks, such as image classification, object detection, and semantic segmentation
Brainchip says that these devices are extremely energy-efficient in running complex models. For instance, it can run RESNET-50 on the neural processor. It is capable of spatial-temporal convolutions for handling 3D data. It enables advanced video analytics and predictive analysis of raw time series data. It allows low-power support for vision transformation for edge AIoT.
The akida product is being offered in three types of packages. The akida-E is the smallest and lowest power with 1-4 nodes and is designed for sensor inference and is used for detection and classification. The Akida-S with 2-8 nodes, includes a microprocessor with sensor fusion and includes application system on chips (SoCs) and is used for detection and classification working with a system CPU. The akida-P with 8-128 nodes, is the company’s maximum performance package. It is designed for network edge inference and neural network accelerators and can be used for classification, detection, segmentation and prediction with hardware accelerators.
Brainchip believes its akida IP packages can serve many applications including industrial uses such as predictive maintenance, manufacturing management, robotics and automation and security management. In vehicles it can enhance the in-cabin experience, provide real-time sensing, enhance the electronic control unit (ECU) and provide an intuitive human machine interface (HMI). For health and wellness applications it can provide vital-signs prediction, sensory augmentation, chronic disease monitoring and fitness and training augmentation. For home consumers it can augment security and surveillance, intelligent home automation, personalization and privacy and provide proactive maintenance.
Hailo announced its Hailo-15 family of high-performance vision processors, designed for integration into intelligent cameras for video processing and analytics at the edge. The image below shows an image of the Hailo VPU package. The company says that Hailo-15 allows smart city operators can more quickly detect and respond to incidents; manufacturers can increase productivity and machine uptime; retailers can protect supply chains and improve customer satisfaction; and transportation authorities can recognize everything from lost children, to accidents, to misplaced luggage.
The Hailo-15 visual processing unit (VPU) family includes three variants — the Hailo-15H, Hailo-15M, and Hailo-15L. Hailo devices also include neural networking cores. These devices are designed to meet the varying processing needs and price points of smart camera makers and AI application providers. The product support from 7 TOPS (Tera Operation per Second) up to 20 TOPS. The company says that all Hailo-15 VPUs support multiple input streams at 4K resolution and combine a powerful CPU and DSP subsystems with Hailo’s AI core. The Hailo-15H is capable of running the object detection model YoloV5M6 with high input resolution (1280×1280) at real time sensor rate, or the industry classification model benchmark, ResNet-50, at an 700 FPS.
the Hailo-15H is capable of running the state-of-the-art object detection model YoloV5M6 with high input resolution (1280×1280) at real time sensor rate, or the industry classification model benchmark, ResNet-50, at an extraordinary 700 FPS. The figure below shows the block diagram for the Hailo device.
This is an industrial SoC with interfaces for image sensors, data and memory. It utilizes a Yoto-based Linux distribution and provides secure boot and debug with a hardware accelerated crypto library including TrustZone, and a true random number generation (TRNG) and Firewall. Hailo says that the low power consumption of these devices enables implementation without an active cooling system (e.g. a fan), making it useful in industrial and outdoor applications.
AI provides ways to process the vast amounts of stored and generated data by creating models and running them on inference engines in devices and at the network edge. Brainchip and Hailo recently announced neural network-based AI inference devices for industrial, medical, consumer, smart cities and other applications that could augment human abilities.