Home monitoring: How to choose memory technology to adapt to new system architectures

In recent years, the demand for products using artificial intelligence (AI) and machine learning (ML) technologies has surged. However, studies have found that it is extremely difficult to make machines learn the skills and judgment that humans are born with in applications such as autonomous driving. Although the hype about AI has exceeded reality in some areas, there are still many real products using ML functions that are gradually attracting consumers. For example, security and home monitoring systems using intelligent vision have huge potential: analysis firm Strategy Analytics predicts that the home security camera market will grow by more than 50% between 2019 and 2023, and the market value will increase from US$8 billion to US$13 billion.

Smart cameras have flourished as image and scene recognition are considered to be among the most suitable functions for the development of ML technology. Smart features in home monitoring systems can be used to:

• Caring for the elderly or vulnerable

• Monitor whether the baby's breathing is normal during sleep

• Recognize the face of a resident or pet (e.g. smart doorbell or smart pet door) to automatically allow them in

• Monitor suspicious activity around your home and trigger intrusion alarms

These new smart surveillance systems, equipped with advanced image signal processors (ISPs), are essentially computers with specialized functions. The latest products in this category use a computer-like architecture that relies on the low-latency, high-bandwidth operation of high-speed DRAM systems to store the application code that executes on the ISP. This article will explore the factors that influence the selection of DRAM technology in the new category of home security surveillance systems and explain how DRAM developments address the specific needs of ISP-type architectures.

Fully functional computer camera system memory

The latest smart cameras for home security monitoring use neural networks for image or scene recognition. These monitoring systems include an inference engine: a software algorithm that has been trained to “learn” to recognize image or scene types by viewing and analyzing thousands or millions of labeled images contained in a training data set.

Figure 1: NetGear Arlo, one of the early AI home security cameras. (Image credit: Scott Lewis, licensed under Creative Commons)

Executing an inference engine is a computationally intensive workload, so some of the latest home monitoring systems are equipped with high-performance integrated ISPs and AI inference engines from chipset manufacturers such as Ambarella. These ISPs are usually based on powerful Arm® Cortex®-A series application processor cores.

Another common approach to home security camera design is "Light AI," which uses smaller, less-functional chipsets from manufacturers such as Omnivision, Kneron, and NXP Semiconductors. While Omnivision and NXP are already world-renowned suppliers of advanced sensor and processor chips, Kneron has rapidly risen in recent years and has become a leader in AI technology, with its KL520 being selected by EE Times as one of the top ten chipsets for edge AI applications.

Therefore, these AI-based home surveillance systems are actually computer cameras (see Figure 1), which are completely different from old-fashioned closed-circuit television (CCTV) cameras or security cameras that simply record scenes with time-stamped video clips. The new generation of home security cameras uses a computer-type architecture and is equipped with sufficient high-speed DRAM memory to store complex inference engines and other application code (see Figure 2).

Figure 2: A typical system architecture for a home security camera includes DRAM system memory with ISP support (Ambarella CV25S as an example)

However, unlike tablets, smartphones, or other general-purpose computers, home surveillance systems have dedicated imaging and reasoning functions that fully utilize the performance of components and system architecture. This means that DRAM is purely for image signal processing functions rather than general computing tasks.

As a result, home security camera manufacturers have found that specialized low- or medium-density LPDRAM has advantages over using the high-density 4GB or 8GB DRAM commonly found in smartphones or laptops.

Figure 3: The Kneron KL520 SoC is on the display board. The KL520 neural processing unit (NPU) was selected as one of the top ten chipsets for edge AI applications by EE Times (KL520 contains Winbond 512Mb LPDDR2 KGD)

Energy-saving, long-lasting and reliable

Energy consumption issues have a significant impact on manufacturers' choice of DRAM type and DRAM supplier. For high-end computers and tablets, OEMs need to maximize DRAM memory capacity and bandwidth while minimizing costs, so they must choose the latest DRAM technology using leading-edge process technology.

In contrast, home security and surveillance cameras focus more on:

• Power conservation: For example, a smart video doorbell must be designed to extend the operating time of its disposable alkaline batteries to minimize the frequency of battery replacement.

• Longevity: The economics of the home surveillance system market dictate that a successful design must remain in the market for more than three years. Therefore, suppliers must ensure that the components in these designs have a long lifespan to avoid OEMs having to regularly redesign the hardware.

• Reliability: Since home surveillance systems must be installed in precise fixed locations to capture the desired scenes, repairs and replacements require on-site visits by technicians, which is much more expensive than sending a faulty laptop back to a repair center. Therefore, both customers and OEMs attach great importance to the reliability of individual components and the overall system.

The above factors will also affect OEMs' evaluation of DRAM products and suppliers. The approaches of dedicated memory manufacturers and large DRAM manufacturers that use cutting-edge processes and focus on the PC and server markets are in stark contrast.

DRAM specialists offer a wide range of products for applications that can be run with conventional DRAM technology, rather than offering state-of-the-art products. For example, Winbond offers standard versions of SDRAM, including DDR, DDR2 and DDR3, as well as low-power LPDDR, LPDDR2, LPDDR3, LPDDR4 and LPDDR4x mobile DRAM for power-sensitive designs.

The processes of mainstream DRAM manufacturers are also optimized for memory capacity and die cost. Manufacturers who choose not to adopt cutting-edge technology are free to improve traditional processes to more fully meet the needs of industrial, automotive, medical and consumer markets. The processes implemented by Winbond's mobile memory wafer fab in Taiwan can minimize the power consumption of the device, helping Winbond's standard version DRAM products provide better power performance than the standard products of competitors.

Home monitoring system OEMs also need to rely on the reliability of their selected components, and dedicated memory suppliers are also configured to meet market needs. Winbond's LPDRAM can provide high reliability guarantees.

Finally, while mainstream high-capacity DRAMs are interface-compatible with the x86 microprocessor architecture used in PCs and servers, specialized DRAMs can be adapted to integrate more easily with other types of processors, especially Arm architecture devices. To simplify system integration for home surveillance system manufacturers, Winbond has partnered with ISP chipset manufacturers to support reference design boards for smart surveillance cameras.

Supports rapid development of more advanced monitoring systems

As AI and machine learning technologies advance, the market demand for greater signal processing and higher-speed memory will continue to grow to support ISPs. The development of dedicated DRAM should include the introduction of new standard technologies (such as LPDDR4/4x) to provide greater bandwidth. Winbond's DRAM and Flash memory products are backed by international quality certifications and high-quality manufacturing processes, giving OEMs the confidence to build smarter new features in home and security monitoring systems.