Definition of Computing Platforms
A computing platform refers to a complete environment that supports the operation and development of applications, consisting of underlying hardware architecture and software frameworks. Its core is to provide a foundational support environment for computing tasks through an integrated soft and hardware environment. The characteristics include providing a stable operational foundation and development interface for upper-layer software through a standardized technology stack (including processors, operating systems, runtime libraries, development tools, and key services).
From a technical composition perspective, computing platforms can be divided into hardware and software layers: the hardware layer includes physical hardware such as Central Processing Units (CPUs), Graphics Processing Units (GPUs), and memory; the software layer encompasses operating systems, virtual machines, containers, programming frameworks, and Application Programming Interfaces (APIs). These components together build a scalable ecosystem that enables developers to create compatible applications based on unified standards.
Modern computing platforms have further evolved into service-oriented ecosystems, providing more efficient resource scheduling and development tools and operational environments through hardware abstraction and resource virtualization. For example, Platform as a Service (PaaS) and artificial intelligence development platforms (such as TensorFlow and PyTorch) further shield the complexity of underlying infrastructure, achieving dynamic resource scheduling and cross-environment deployment through containerization, microservices, and DevOps practices.
Edge artificial intelligence computing platforms are deployed in dedicated computing environments at the network edge, close to the data generation source. They integrate specialized hardware architectures (such as NPUs, GPUs, ASICs, and other acceleration units) aimed at AI processing with optimized software stacks (including lightweight inference frameworks, model management tools, and edge operating systems) to provide localized computing services for AI applications that are low-latency, energy-efficient, highly reliable, and capable of privacy protection.
Types of Edge Artificial Intelligence Computing Platforms
Edge artificial intelligence computing platforms can be classified into four types based on their deployment location, computing capabilities, and core functions: embedded platforms for terminal devices, edge gateway platforms, edge server platforms, and edge-cloud collaborative platforms.
Embedded Platforms for Terminal Devices. These platforms are integrated within terminal devices and serve as the most peripheral carriers of AI computing power. Their main characteristics are high integration and ultra-low power consumption, typically using system-on-chip (SoC) or microcontroller (MCU) designs that integrate CPUs with dedicated AI acceleration cores, optimized for executing pre-trained single lightweight inference tasks in environments with extremely limited resources (power, computing power, size), thus enabling the intelligence of terminal devices. Examples include facial recognition in smart locks and predictive maintenance.
Edge Gateway Platforms. These platforms act as hub nodes connecting terminals to the cloud, equipped with richer interfaces and moderate computing power. They can aggregate data from multiple terminals and run more complex AI models, making local decisions and preprocessing data to reduce the burden on the cloud. The hardware form is typically based on high-performance AI chips in industrial intelligent gateways or computing boxes, such as in factory production line quality inspection scenarios.
Edge Server Platforms. These provide high-performance AI computing capabilities close to the cloud. Such platforms are usually deployed in regional centers or network sides, forming a micro data center at the edge. They can handle extremely complex AI inference tasks and support local data for model fine-tuning and continuous learning, such as in local factory analysis and autonomous driving scenarios.
Edge-Cloud Collaborative Platforms. These platforms essentially serve as a software-defined management and control interface. By abstracting and scheduling widely distributed edge hardware resources through a unified cloud-native technology stack, they achieve full lifecycle automated operations for AI tasks from issuance, deployment to version management, thus building a globally collaborative and efficiently managed “edge as a service” system that supports large-scale intelligent application networks across regions. These platform types collaborate to form a complete and efficient layer of edge intelligent infrastructure.
Types of Processors in Edge Artificial Intelligence Computing Platforms
The types of processor products in edge artificial intelligence computing platforms mainly include the following:
Central Processing Unit (CPU), as the core of the computing system, is a processor that executes software instructions and handles general computing tasks. It is known for its powerful logical control, complex serial processing capabilities, and excellent versatility, capable of running complex operating systems (such as Linux and Windows) and undertaking various diverse computing tasks. In edge AI systems, the CPU typically plays the role of a “manager” and “coordinator”: responsible for system control, task scheduling, data handling, and running algorithm parts that are not suitable for execution on other accelerators, such as x86 and ARM Cortex-A series.
Microcontroller Unit (MCU), is a miniature computer that integrates a central processing unit, memory, timers/counters, and various input/output interfaces on a single integrated circuit chip. Microcontrollers are the foundational processors for terminal embedded devices, enabling functional computing tasks. Due to their limited resources, they can run lightweight machine learning. With the introduction of neural network processors (NPUs) as accelerators, microcontrollers have significantly enhanced their processing capabilities in artificial intelligence, further expanding their application range in edge computing.
Micro Processor Unit (MPU), this concept is sometimes used interchangeably with CPU in the industry, referring to a programmable general computing core that integrates the functions of a computer’s central processing unit (CPU) onto one or several integrated circuit chips. It can be viewed as a more powerful CPU chip designed for running advanced operating systems and complex applications. On the edge, MPUs provide a strong local computing foundation for devices.
Graphics Processing Unit (GPU), originally designed for graphics rendering, has become the mainstream acceleration solution for AI training and inference due to its powerful parallel floating-point computing capabilities. On the edge, GPUs can efficiently process unstructured data such as images and videos, executing complex deep learning model inference tasks.
Field-Programmable Gate Array (FPGA), provides hardware-level reconfigurability, suitable for scenarios requiring rapid algorithm iteration or customization. Users can program its internal logic units and connections using hardware description languages (HDL) to build dedicated hardware acceleration circuits, achieving low-latency and high-deterministic parallel computing. FPGAs have unique advantages in fields requiring flexible adaptation, such as industrial automation and communication base stations.
System on Chip (SoC), is the mainstream form of modern edge AI processors. It integrates various computing units (such as CPUs, GPUs, NPUs, DSPs, etc.) along with memory controllers, various I/O interfaces, and peripheral controllers onto a single chip. SoCs achieve optimal performance, power consumption, cost, and size balance through heterogeneous computing architectures, allowing tasks to be executed on the most suitable processing units.
Application-Specific Integrated Circuit (ASIC), is a chip designed specifically for a particular purpose or algorithm (such as neural network inference).
Neural-network Processing Unit (NPU), is a type of ASIC designed specifically to accelerate neural network computations (such as convolution, pooling, activation functions, etc.). NPUs are typically integrated into SoCs as IP cores rather than standalone chips, providing efficient and low-power AI inference capabilities for systems.
Edge Artificial Intelligence Development Platforms
There are several types of edge artificial intelligence development platforms:
ST Edge AI Suite, includes several core tools such as STM32Cube.AI, NanoEdge AI Studio, MEMS Studio, and Stellar Studio. STM32Cube.AI is one of the core development tools provided by STMicroelectronics, which converts pre-trained AI models into optimized C code, supporting mainstream AI frameworks like TensorFlow Lite and ONNX. NanoEdge AI Studio is a low-code/no-code development tool based on AutoML technology that can automatically generate lightweight machine learning libraries with minimal data and supports on-device learning. ST Edge AI Developer Cloud is an online collaborative platform that allows developers to optimize and benchmark edge AI models online. The AI for OpenSTLinux (X-LINUX-AI) is an embedded Linux system development kit for STM32 MPUs, supporting the deployment of AI models on STM32 MPUs to meet high scalability application needs, integrating Linux AI frameworks and application examples, and supporting the processing and deployment of complex models such as time series, audio, and vision.
Edge Impulse, is a cloud platform that provides users with low-code or even automated machine learning development experiences. It greatly simplifies the entire process from data collection to model deployment, allowing users to easily complete data preprocessing, model training, and performance testing through a graphical interface, ultimately generating optimized models that can run directly on embedded devices. This platform is particularly suitable for rapid AI prototyping and deployment in Internet of Things (IoT) devices and embedded systems.
SensiML, is an end-to-end platform focused on simplifying TinyML code development. Its core advantage lies in its ability to automate the generation of efficient AI algorithms and corresponding firmware code. The platform provides a series of tools for data collection, automatic labeling, model training, and code generation, aimed at enabling developers without a deep machine learning background to develop applications for resource-constrained MCUs, such as predictive maintenance, anomaly detection, and gesture recognition.
NVIDIA Jetson, is a powerful ecosystem that integrates high-performance hardware with a complete software stack. This platform offers hardware computing modules ranging from entry-level to ultra-high performance (such as the Jetson Orin series), all equipped with GPUs and providing strong AI computing power. Its software stack (JetPack SDK) includes tools like CUDA, cuDNN, and TensorRT, providing developers with a familiar development environment, making it an ideal choice for high-performance edge AI applications in robotics, autonomous driving, and smart cities.
Google Coral, is a complete edge AI solution launched by Google, which includes both hardware and software. Its core is the Edge TPU ASIC chip designed for TensorFlow Lite model inference, providing high-efficiency AI acceleration. The platform offers development boards, USB accelerator sticks, and corresponding model conversion and deployment tools, primarily targeting applications in visual processing, IoT, and robotics.
AWS IoT Greengrass, is the edge runtime software launched by Amazon Web Services (AWS), which seamlessly extends AWS cloud service capabilities to local devices. Developers can run AWS Lambda functions and pre-trained machine learning models for local inference on edge devices while enjoying the powerful device management, security, and data analysis services provided by AWS cloud. It is particularly suitable for large-scale industrial automation and smart device projects that require tight integration with the AWS ecosystem.
Azure IoT Edge, is the edge computing component of Microsoft Azure cloud services, allowing cloud-based AI analytics and custom business logic to be containerized and deployed to run on edge devices. It supports offline task execution in disconnected environments and can be centrally managed and monitored through Azure cloud, widely used in scenarios such as predictive maintenance and real-time data analysis.
Qeexo AutoML, is a platform focused on providing automated machine learning solutions for ultra-low-power microcontrollers (MCUs). Its feature is a one-click automated ML workflow, particularly adept at handling time-series data from various sensors and generating extremely compact and efficient code, making it very suitable for application development in industrial, IoT, and wearable device fields.
Imagimob Studio, is a platform designed for rapid development of end-to-end Edge AI applications on resource-constrained devices. It provides an integrated development environment that guides users through the entire process from data preparation, model design, training to deployment, with wide applications in audio classification, predictive maintenance, and gesture recognition.
Fibocom AI Stack, is the overall edge AI solution launched by Fibocom, which integrates high-performance communication modules, AI toolchains, and a vast array of optimized models. This platform provides model compression and optimization, high-performance inference engines, and deployment capabilities across different chip platforms, aiming to promote the implementation of AI applications in fields such as smart retail, autonomous driving, and wearable devices.
NXP eIQ® ML, is the machine learning software development environment provided by NXP Semiconductors for its microcontrollers (MCUs) and application processors (MPUs). It includes the eIQ Toolkit toolchain, various inference engines (such as GLOW, ARM NN), neural network compilers, and optimization libraries, helping developers efficiently implement functions such as object detection and voice commands on their NXP hardware.
Stream Analyze, is a platform focused on providing end-to-end development capabilities for edge devices in industrial and automotive fields. It is designed for processing real-time data streams, supporting the development, training, deployment, and coordination of analysis, computation, and AI models, and has been applied in complex scenarios such as mining loaders and autonomous vehicles.
Wangsu Technology Edge AI Platform, is a full-link platform for enterprise applications, providing comprehensive capabilities covering “model access – inference optimization – scenario implementation.” Its core products include edge AI gateways, edge model inference services, and edge AI applications, aimed at supporting the large-scale implementation of enterprise-level AIGC applications such as intelligent customer service and content generation (AIGC).
Conclusion
The edge artificial intelligence ecosystem is driven by processor hardware and development tool software. The hardware layer presents a diversified architecture: general-purpose CPUs are responsible for system control, GPUs accelerate parallel computing, dedicated ASICs/NPUs focus on efficient inference, FPGAs provide flexible acceleration, and MCUs meet low-power requirements. Modern SoCs adopt heterogeneous computing architectures, integrating multiple processing units to optimize performance-to-power ratios. The software layer forms a full-stack development ecosystem, covering the entire toolchain from model optimization to deployment, with cloud service providers offering standardized cloud-edge collaborative solutions. Through automated optimization tools and pre-integrated model libraries, the development threshold is significantly lowered, promoting the large-scale commercial deployment of AI technology in fields such as industrial automation and the Internet of Things.