The edge artificial intelligence (AI) chipset market is expected to exceed the cloud AI chipset market for the first time in 2025. According to global tech market advisory firm, ABI Research, the edge AI chipset market will reach US$12.2 billion in revenues, outpacing the cloud AI chipset market, which will reach US$11.9 billion in 2025. Edge AI includes smart phones, wearable devices, smart vehicles, and smart home/industry/city. Above of them, smart home will play a key driver role of edge AI market.
Most AI training workloads happen in the public and private clouds. Traditionally, the centralization of these workloads in the cloud brings the benefits of flexibility and scalability. However, driven by the need for privacy, cybersecurity, and low latency, performing AI inference workloads is preferred on gateways, devices, and sensors. Recent advancements in key domains, including new AI learning architecture and high-performance computational chipsets, have played a critical role in this shift.
Edge AI computes the data as close as possible to the physical system. The advantage is that the processing of data does not require a connected network. The computation of data happens near the edge of a network, where the data is being developed, instead of in a centralized data-processing center. One of the biggest benefits of edge AI is the ability to secure real-time results for time-sensitive needs. In many scenarios, sensor data can be collected, analyzed and communicated straightaway, without having to send the data to a time-sensitive cloud center.
In addition to applying in smart phones, deep learning is also used in IoT devices (also known as AIoT). Edge AI bring a new concept for legacy IoTs. Nevertheless, the computing power of MCU, which is a traditional processing unit in IoTs, is too weak to doing deep learning. At this moment, there are two kind of hardwired mechanism to help MCU: DSP or the dedicated accelerator (called as Deep Neural Network/DNN). These hardwired mechanism could be implemented as IP or a chip.
The fundamental component of both the convolution and fully-connected layers, which are main algorithms of DNN, are the multiply-and-accumulate (MAC) operations. In order to achieve high performance, highly-parallel computing methods are frequently used. Many iterations of weight which are commonly stored in DRAM are updated according to different training styles.
Regardless of above mechanism, DRAM throughput is key to DNN. Therefore, choosing a proper DRAM is critical to AIoT application. Unlike cloud environment, low power is a concern in edge computing. The planner of AIoT device must seek to balance performance and low power. Apart from power and performance, the density of DRAM integrated in AIoTs devices usually only need low density (1~2Gb) rather than commodity one (8~16Gb per die).
Winbond Electronic Corp.’s 1Gb LPDDR3 DRAM die was one such example with AI company Kneron having selected it for its latest system-on-chip (SoC), the KL720. It’s one of several SOCs the company offers that being used in a variety of edge devices, including battery-powered applications such as smart locks and drones that take advantage of a 512Mb LPDDR2 from Winbond.
LPDDR3, delivering a maximum bandwidth 8.5GB/s with a dual 1.2V/1.8V supply, enable customer devices like Kneron’s KL720 to process 4K, Full HD or 3D sensor video images in real time to support AI applications such as face recognition in security cameras or gesture control in public kiosks, as well as perform natural language processing.
Beyond the needs of Kneron and the applications for its KL720 SoC, there are potential uses for devices with the density and bandwidth of Winbond’s LPDDR3 DRAM in automotive applications such as Advanced Driver Assistance Systems (ADAS), which employ cameras that must process video images in real time. Meanwhile, there are many opportunities for IoT endpoints that need to do basic AI inference, which require low densities, but high bandwidth.
LPDDR4/4x x32 has an almost double throughput than LPDDR3, same advantages as what LPDDR3 has with LPDDR2, shown in Fig.1. In power consumption, the IO voltage of LPDDR4x is 0.6(V), in contrast that LPDDR4 is 1.1(V). Although JEDEC has already published the latest LPDDR5 standard, the density of LPDDR5 just released into the market is still too high to be applied in the AIoTs. At this moment, LPDDR4x is still the best choice, if we need more AI computing power than what LPDDR3 could provide.

Fig.1
Winbond owns self-built wafer fabs, and is one of the top four IC makers that can simultaneously provide DRAM and NOR/NAND Flash. The capacity of Winbond LPDDR4/4x DRAM series is 1~4Gb with 25nm technology node which is self-developed by Winbond, and the speed could be up to 4266Mbps. In addition to the supply type of known good die, Winbond also provides the standard 200BGA package.
-
With the evolution of automobiles to digitalization, the demand for data storage and the transmission of information is getting higher. All automotive applications require qualified storage products and devices that will endure in embedded environments exposed to extreme temperatures...
Flash Memory Market Ushered in Fierce Competition with the Digitalization of Electric Vehicles
-
According to PwC's estimate, the metaverse market will exceed $1.5 trillion by 2030, with semiconductors being one of the key driving factors in its development. The semiconductor industry will likely usher in new growth with the progress of the metaverse market, with the strongest push coming from semiconductor products related to computing.
The New Role of Storage Manufacturers in the Metaverse
-
AI technology is divided into two categories; Training and Inference. AI-related chips include CPU, GPU, FPGA, TPU and, ASIC. To get an idea of how these chips compare to one another, here is a comparison focusing of 5 key factors...
Choosing the Right Flash Memory for AI Endpoint Applications
-
ISO21434 has been made mandatory by many car makers and their component suppliers, starting from mid-2022. As a result, the automotive industry is now required to significantly improve how cyber threats are managed. As this standard applies to both the modules and their components, it requires the automotive industry to adapt devices capable of meeting this standard and provide the required protection against cyber threats.
Achieving ISO/SAE21434 Cyber Security Using Secure Flash
-
The past 18 months have seen unprecedented market change brought on by the global pandemic - which, like dropping a stone in a pond, will continue to create expanding ripples of change throughout the industry. One such ripple has been the global shortage of ICs, affecting the electronics industry and the worldwide economy.
Memory without Compromise
-
Leveraging Octal interfaces for NAND flash memory will enable automotive, AI, consumer electronics, and industrial manufacturers to tap into this market growth opportunity by providing code storage in high density without having to pay a premium for NOR flash, a fast memory technology which scales poorly at densities above 512Mbits.
The Rise of OctalNAND for Automotive, OTA, AI and More!
-
Automobile electronic systems are steadily becoming more intelligent. Advanced electronic functionality is being added throughout the vehicle such as ADAS, Gateway, Power Train, Infotainment, V2V, and V2X. These new capabilities are driving the need for increased security and safety.
The Hidden Security Risks of Automotive Electronic Systems
-
The growth in the number of processor-based Electronic Control Units (ECUs) in cars is a familiar trend. Of equal significance to automotive system architects is the concurrent growth in the code footprint of new and developing applications in the car, a result in part of the increasing...
QspiNAND with ultra-fast write speed: A new option for over-the-air updating of automotive code
-
Owing to the fact COVID-19 blocks transportation, the process of penetration is lower than we expected. Most of IoT applications are still connected through 2G or 3G technology. In 2019, the estimated number of massive IoT applications have grown to triple and will reach approximately 100 million by the end of 2020.
HyperRAM™- Best DRAM choice for IoT application
-
A new generation of secure Flash memory products has come on to the market to provide a secure hardware foundation for embedded devices which do not require payment-grade protection. Often featuring a standard Flash memory package footprint and pin-out, and controlled via the standard SPI NOR Flash instruction set, these secure Flash memories are easy for general embedded device designers to implement a comprehensive set of security functions to protect connected devices from attack on a system’s integrity or data privacy.
How new secure flash devices promise security for IoT devices’ code and data
-
New intelligent vision systems for the home, based on advanced image signal processors (ISPs), are in effect function-specific computers. The latest products in this category have adopted computer-like architectures which depend for low latency, highly responsive operation on fast DRAM system memory to store the application code running on the ISP...
How the architecture of new home security vision systems affects their choice of system memory technology
-
As the world emerges from the Covid-19 crisis, it is likely to leave lasting impacts not only on the patients who have been treated by dedicated doctors and nurses, but also generally on the way that medicine is practised. As a leading manufacturer of specialty memory ICs, Winbond is constantly looking ahead to understand the broad trends in areas such as the medical equipment market...
New generation of wearable medical devices calls for secure, high-density non-volatile memory