i卡會員

The booming development of artificial intelligence (AI) is drawing various industries around the world into the AIoT ecosystem. Many companies have been actively constructing their roles or positioning themselves in the AIoT ecosystem in order to create business opportunities in AIoT applications by working together.     

Taiwanese manufacturers have a long track record in the manufacturing of ICT products for international brands and have built a reputation for quality and cost advantages in hardware. However, the mega trends in artificial intelligence, IoT (Internet-of-Things) and Big Data present a new set of challenges and requirements in cloud computing efficiency, AI algorithm capabilities and Big Data analytics. Among these, AI edge computing is currently the focal point of global industry attention.

The uploading of vast amounts of IoT data to the cloud for AI computing, training and analysis is likely to compromise the efficiency of computing, analytics, data response and machine operation in general. Therefore, it is necessary to shift a portion of the machine learning mechanism from the cloud to the terminal devices. In other words, the data training, analysis and forecasting that was previously processed on cloud servers or in data centres is now going to be handled on the edge nodes of the logical network. Therefore, it is a prerequisite to enhance the data processing capability of the edge nodes.

According to the forecast from IDC, a market research firm, the market for edge computing in terms of terminal and network equipment will exceed $50 billion in 2020. By then, over 50% of data will be analyzed, processed and stored on network edges.

The Industrial Technology Research Institute observed that in 2017 and 2018, industry players in telecommunication, networking equipment, servers and platforms, system integration, semiconductor chips, components and terminal devices have been proactively developing a diversity of solutions and products for edge computing.

The trend in AI edge computing is moving from connectivity across networks in proximity and within the same regions at the hub level towards applications and scenarios of edge devices at the terminal level. In other words, some of the machine learning techniques and autonomous decision-making mechanisms in artificial intelligence will be transferred from the cloud to the edge.

This trend brings three major developments:

First, on-line training is gaining traction and on-device training will become important. Edge computing is about moving the computing for apps, data and services to edge nodes in the network architecture. From the perspective of flexible deployment, real-time response at the user end, bandwidth and cost considerations, edge computing will lead to better efficacy and real-time experience. As demonstrated by the solutions of some leading companies, IoT devices do not have to stay connected all the time. This approach resolves issues of bandwidth, power and computing resources. In other words, AI edge computing is enabled with both on-line training and on-device training. Training is conducted when connected for data transmission. Meanwhile, IoT devices should be equipped with certain decision-making and real-time response capabilities to function while offline.   

Second, edge computing utilizes edge AI chips combined with AI algorithms on a global scale. For instance, Amazon acquired the chip design company Annapurna Labs and has been recruiting chip design talent for the purpose of shifting some of its  smart assistant Alexa’s functions from the cloud to the Echo smart speaker device. Microsoft is working with Intel’s FPGA chips to develop Project Brainwave, a hardware platform. It is also collaborating with MediaTek for MT3620, the system-on-chips that support Azure Sphere. This microcontroller enables IoT devices with embedded security and connectivity functions. Google is developing Cloud IoT Edge and Edge TPU chips, so that edge devices can perform machine learning and inferences. In sum, leading companies are integrating chip architectures by taking into account the necessary platforms, systems and software. The combination of chips and algorithms will define the direction of hardware and software for edge computing going forward.

Third, the trend is for the ecosystem of applications oriented at AI and machine learning.

The ecosystems of edge computing consist of the cross-device connectivity, driven by telecommunication companies, and the variety of applications created by platform heavyweights. At this juncture, the ecosystem for machine learning, enabled by semiconductor chips at the bottom layer, is also taking shape. For example, the semiconductor company ARM from the UK offers solutions such as CPU, GPU, NPU and opensource software to support the global development of the AI edge computing ecosystem. This is expected to benefit Taiwanese players who focus on the hardware for smart edge devices in the pursuit of opportunities in the market for AI edge computing.  

From telcos, IoT device and equipment suppliers, cloud platforms and system integrators to semiconductor companies, many players are actively deploying robust solutions in the AI edge computing space. Taiwanese companies may explore opportunities in AI edge computing by lowering deployment costs and enabling computing resources, flexibility, scalability, real-time responses and usability.

Given Taiwan’s strength in the hardware for smart devices, on-device training should be the preferred route for Taiwanese companies. It is suggested that manufacturers in Taiwan focus on terminal devices by simplifying the product design, optimizing the computing requirements, filtering effective data or converging shared functionality. This will bring Taiwanese companies to different regional markets by circumventing the dominance of cloud platforms and 5G companies.