Advantages of Using Serial NAND Flash as Storage for AI Systems in Embedded System Development

For the innovative potential brought by artificial intelligence (AI), embedded system developers in various market sectors have shown great interest. However, from the perspective of “innovation,” it seems a bit strange, as the foundational technology of AI itself is not a new concept: the AI system based on IBM’s supercomputer “Deep Blue” defeated world chess champion Garry Kasparov back in 1997.

Nonetheless, in the 20 years following this major breakthrough, the progress of various AI technologies has been quite slow. These technologies were not integrated until the late 2010s, making AI a mainstream technology in embedded system development, and this path has become much smoother due to two main factors: First, the industry can access the vast training databases generated by Internet of Things systems that utilize a large number of sensors; additionally, through popular platforms like YouTube, Instagram, Snapchat, and Facebook Live, the industry has also gained access to a large amount of labeled images, videos, and other forms of data for the first time.

The second factor is the capabilities of the main devices in embedded systems, such as application processors, system-on-chip (SoC), or field-programmable gate arrays (FPGA), which have all reached a critical point where they can act as “inference engines.” This means that these installed systems capable of executing machine learning algorithms can interpret image files and recognize objects such as cats and dogs.

Today, AI technology is expected to be rapidly applied to embedded systems. According to market research firm IDC, by 2023, the market compound annual growth rate for AI processors using edge computing systems will reach 65%. However, this trend of adopting AI technology has also raised questions about whether the program code storage memory used by embedded system developers is appropriate.

Currently, SPI NOR Flash is the preferred memory type for storing embedded system boot code and application code. However, to accommodate the larger code libraries generated by the latest AI applications, SPI NOR Flash faces pressure to provide higher memory storage capacity. Additionally, today’s rich software applications in embedded systems often require periodic online updates to execute security patches and add new features. This periodic update requirement has brought the write speed of SPI NOR Flash into the spotlight.

This article argues that, for the storage of system boot code and application code, embedded system developers should maintain an open-minded attitude towards the potential benefits of replacing SPI NOR Flash with Serial NAND Flash, while also reevaluating existing stereotypes regarding NAND reliability, lifespan, and performance.

Miniaturization Issues with NOR Flash

In the past, SPI NOR Flash was favored in the market for providing reliable storage memory due to NOR Flash’s inherently robust nature and its data retention period of over 100,000 hours.

According to data released by market research firm Web Feet Research in 2018, Winbond Electronics undoubtedly holds the leading position in global SPI NOR Flash manufacturing in terms of output and value, with a product line that covers everything from the smallest 3.0V 512Kb and 1.8V 1Mb to the highest capacity of 2Gb SPI NOR. Winbond’s advanced NOR Flash wafers are currently manufactured using a 58nm process; according to the development path, they will be shrunk to a 45nm process. However, compared to the process shrinkage pace following Moore’s Law, the progress in NOR Flash process technology has significantly slowed in recent years.This slow progress has troubled many embedded system developers who are beginning to adopt AI technology, as the AI applications they use require more complex and larger software program codes.Therefore, developers will increasingly need memory storage capacities of over 1Gb, and within this storage capacity range, Serial NAND Flash clearly comes out on top in terms of cost per bit.

Unlike the SPI NOR Flash product line, NAND Flash generally follows Moore’s Law in process node shrinkage, shrinking from 46nm, 32nm, 2x nm all the way down to 1x few nm. Recently, the development of 3D NAND structures has also allowed NAND Flash manufacturers to continue increasing product memory density after the 1x nm process node. Since, in the semiconductor industry, die area is closely related to product cost, if the process node used to manufacture Serial NAND continues to shrink, its cost in the high-density storage field of over 1Gb will be significantly lower than that of SPI NOR products.

Furthermore, today’s smart connected devices mostly have over-the-air (OTA) firmware update capabilities to perform security fixes or function upgrades. In this regard, SPI NOR is also at a disadvantage. In a typical example of OTA update operation, the latest code overwrites the backup code in the flash array so that the latest code will be executed on the next boot. Therefore, in this case, the key parameters for update performance are write time and erase time; in this regard, Serial NAND also surpasses SPI NOR memory.

Evaluation Issues with NAND Flash

Despite the cost and performance advantages of Serial NAND Flash in software-rich embedded AI applications, the embedded system community still needs to change its mindset and be willing to view Serial NAND as a storage solution for boot and application code. They have long limited the use of NAND solely to ultra-high-capacity data storage, leading to preconceived notions about Serial NAND Flash.

For solid-state drives (SSD) in laptops or tablets, manufacturers sacrifice data integrity and data retention years to achieve high storage capacity and low cost per bit by adopting the latest 3D multi-level cell technology. In the real world, it is an acceptable compromise to let users’ music or video files experience some bit corruption or loss over time in order to use the latest 1x process node to manufacture ultra-low-cost memory.

However, for storing system boot code or application code, using the highly stable SLC architecture of Serial NAND Flash is the best choice, as each has its strengths compared to advanced/ultra-low-cost 3D NAND.

Serial NAND emphasizes performance and reliability. To provide a transition path from SPI NOR to Serial NAND Flash in AI application storage, Winbond has perfected its Serial NAND process and serial interface to offer:

Fast read performance
Fast write/erase performance
Interface compatibility equivalent to SPI NOR
High data integrity and long-term data retention

The high reliability of Winbond’s QspiNAND (Quad SPI NAND) Flash comes from its process:Using 46nm process manufactured single-level cell (SLC) Flash, although it is about three generations behind the 3D multi-level cell (MLC) Flash used for consumer SSDs, the 46nm process provides sufficient charge storage space, ensuring a longer data retention period:Winbond’s QspiNAND Flash can retain data for up to 10 years within its rated operating temperature range.

Due to the 46nm process NAND architecture, the die area per bit and cost per bit of Serial NAND are significantly lower than those of the 58nm process SPI NOR, which is the main reason why Serial NAND has a cost advantage of up to half in products over 1Gb capacity compared to SPI NOR.

Additionally, the built-in error correction code (ECC) circuit can maintain data integrity during write and read operations, which is also a significant advantage. Winbond’s QspiNAND Flash requires only 1-bit ECC for maintenance but provides 4-bit EC storage space, thus easily supporting any SoC, application processor, or other advanced computing platforms.

Second Generation Serial NAND Performance Evolution

The latest developments in Winbond’s interface technology have enabled the latest generation of Serial NAND Flash devices to outperform SPI NOR products in both performance and cost competitiveness in embedded AI-oriented applications.

Advantages of Using Serial NAND Flash as Storage for AI Systems in Embedded System Development

Figure 1: In the past, SPI NOR provided faster read speeds than Serial NAND (Image Source: Winbond Electronics)

Latency is a key performance parameter for AI systems; inference engines using machine learning algorithms must perform complex calculations within milliseconds, so fast data read performance is required.

Winbond’s first-generation 104MHz QspiNAND Flash achieves a maximum data throughput of 50Mb per second in continuous read mode. In comparison, the 133MHz Quad SpiNOR Flash offers a maximum data throughput of 60Mb per second (see Figure 1).

Winbond has launched the second generation W25Nxx JW QspiNAND series devices, which support higher pulse speeds at 166MHz single-edge (STR) transmission mode and 83Hz double-edge (DTR) transmission mode, with a maximum read throughput of 83MB per second.

By adopting the W72N series Dual QspiNAND Flash products (made from two die in a single package, providing dual x4 memory with 8 I/O channels), this read throughput can be doubled to 166MB per second (see Figure 2).

Figure 2: Winbond’s Dual Quad QspiNAND provides a maximum read speed of 166Mb per second (Image Source: Winbond Electronics).

This rapid read performance is expected to reduce the latency of AI systems. The high performance of Winbond QspiNAND Flash can also support rapid OTA operations, minimizing downtime. The write mode throughput of the latest 1Gb W25N01JW QspiNAND product is 8.5Mb per second, higher than the 0.36Mb per second of Winbond’s SpiNOR Flash product with 256Mb W25H256JV; the erase time for the 128Kb storage block of QspiNAND products is 2ms, better than the 150ms for the 64Kb storage block of SpiNOR products.

The total write time for 1Gb data on Spi NOR devices is 356 seconds, which is nearly 6 minutes; in contrast, the second-generation Qspi NAND products only require 15 seconds.

Easy Integration into Embedded Systems

If OEM manufacturers must integrate AI functionality into embedded systems, storage space must be considered to be over 1Gb. Therefore, evaluating Serial NAND Flash as an expensive SPI NOR alternative becomes a necessary step. Fortunately, Winbond’s QspiNAND series products are all packaged in industrial standard sizes and pin arrangements, so they can directly replace SPI NOR Flash products in existing designs.

Moreover, another powerful boost to integrating this advanced and reliable Serial NAND technology into embedded systems comes from major SoC suppliers, including NXP Semiconductors, STMicroelectronics, and Renesas Electronics. For example, NXP has integrated Winbond’s SpiStack NOR+NAND co-packaged device into the LS1012A edge computing processor and used it on the FRWY-LS1012A development board. In this design, the Winbond QspiNAND device can be used to store the Linux® operating system program code for the development board, while the small Winbond SpiNOR die stores the processor’s boot code.

The second generation QspiNAND Flash has now launched products with storage capacities of 1Gb, expandable to 2Gb, 4Gb, or even higher capacities. As the code libraries continue to grow with the increasingly complex AI edge computing technologies, users can still obtain reliable storage solutions.

If you have any questions, feel free to click on “Read Original” to ask for discussion.

Statement:This article is sponsored by Winbond Electronics, and this public account remains neutral regarding the statements and views expressed in the text.

Related posts:

Leave a Comment Cancel reply