Several vendors are equipping SSDs with artificial intelligence. An AI SSD could help enterprises in such areas as management, performance and security.
AI is attracting a lot of attention, so it's no surprise to find groundbreaking SSDs and controllers adopting it to improve the performance of either the SSD itself or the system that hosts it.
An SSD can incorporate AI in multiple ways, so what is best? Implementations of AI in SSDs that will likely impact the enterprise include managing NAND, screening for malware and pre-processing data.
How AI works in SSDs and top uses
Although most people think of AI as a technology that enables systems to figure out things by themselves, the technology can also sift through large data sets to identify trends.
The internal NAND flash management of SSDs presents that kind of scenario. Wear management, the timing of garbage collection and overprovisioning all involve many tunable factors. SSD designers and SSD controller manufacturers have begun to embrace AI to improve those processes. With those improvements, SSD performance improves as well.
Other applications have started to incorporate AI into the SSD. For example, an AI SSD can offload tasks from the host as a specialized sort of computational storage. That AI can augment an enterprise's security by independently scanning for malware.
Top use #1: Manage NAND
AI can manage hot and cold data on two different kinds of flash chips. This application of AI is a process that many SSDs already perform. SSDs frequently set a block of NAND to operate as single-level cell (SLC) flash, which is faster but several times more expensive than triple-level cell (TLC) or quad-level cell (QLC) flash. Hot data is stored in the SLC blocks and cooler data is moved to the TLC or QLC.
InnoGrit manufactures SSD controllers that integrate AI into the management of hot and cold data. Any SSD maker can use the controller. After a brief training phase, the SSD adapts itself to the workload and manages the placement of hot and cold data.
Another company, DapuStor, uses AI to improve SSD performance by predicting the workload. The company uses the Storage Networking Industry Association's Real World Workloads to train SSDs to perform better in certain environments.
DapuStor's algorithm uses a machine learning approach called long short-term memory (LSTM) to analyze its current workload and determine I/O intensity. Since storage I/O prediction is a time-series problem, LSTM is a good fit. The company claims a prediction accuracy of 95%, resulting in a 20% speed improvement when applied to a commercial SSD.
SSD maker PNY introduced its LX series of SSDs aimed specifically at Chia mining and other applications that require a Proof of Space and Time, an application that creates a high write load. These SSDs combine an advanced AI engine with low-density parity check to improve NAND endurance. The company boasts that its 2 TB model has a "Chia plotting rating" of 54,000 terabytes written.
SSD makers are trying out AI in various forms to improve either the SSD's performance or the performance of the system that hosts the SSD.
Top use #2: Perform malware screening
AI is particularly good at recognizing patterns that might escape more established pattern-matching models.
To this end, Singapore-based SSD maker Flexxon developed what it calls the X-PHY AI Cyber Secure SSD to monitor data streams in real time to watch for malware. This SSD brings an AI coprocessor and firmware into the SSD to create a machine learning system that analyzes low-level storage functions, like read and write patterns. Since ransomware usually follows identifiable data access patterns, Flexxon expects this approach to help detect unknown threats, including zero-day attacks, without any need of malware signatures.
Top use #3: Pre-process data in place
As with computational storage, AI can serve a useful function in offloading data-intensive processes from the server. This means that the server doesn't need to import and export data to and from storage, something that wastes processing power and energy. Instead, that data-intensive function is moved to the data, in this case to the SSD that holds that data.
SSD controller maker Marvell has promoted the idea of using an SSD's internal AI capability to generate metadata for data stored within the SSD. To this end, Marvell and Nvidia have jointly developed a proof-of-concept controller that incorporates Nvidia's Deep Learning Accelerator.
Marvell expects the internal AI capability to generate tags over enormous unstructured data sets, a form of pre-processing that could be useful in vision processing, video analytics and text processing. Marvell's strong position in SSD controllers implies that development of a future SSD based on a Marvell controller is quite likely.
Leading SSD maker Samsung demonstrated an AI SSD proof-of-concept model based on the company's Esperanto neural processor chip. Samsung designed this demonstration to show how it can offload AI tasks into an SSD to lighten the host processor's load.
But wait, there's more
In addition, Synopsys, a chip design IP company, has its DesignWare ARC line of SSD processors. Synopsys says the ARC EV line of processors offers a "fully programmable and scalable solution for AI." Some SSD makers might develop proprietary AI-based SSD controllers.
Challenges and cautions of use
In each of these cases, the inner workings of AI are invisible to the SSD user. Instead, an AI-based function is presented as just that -- a useful function that performs well without user intervention. This has its pros and cons.
To the user's advantage, the function is already built, installed and tested, so enterprises can use it without a lot of effort. The negative side is that it's aimed at a typical scenario that may or may not fit the user's specific application. From that perspective, it's much like using off-the-shelf software.
SSDs with special features of any kind become vendor-specific. Users can't shop from vendor to vendor for the best deal. With a commodity like an SSD, where most users expect to be able to change suppliers to save some money, this acts as a lock-in that prevents the user from changing vendors. As long as users are OK with that situation, there's no reason to avoid an AI SSD.
Most importantly, SSD makers are trying out AI in various forms to improve either the SSD's performance or the performance of the system that hosts the SSD. Over time, expect enterprises to adopt both. It is quite likely that, in five years, nearly all SSDs will include AI in their NAND management in a way so integrated with the drive that vendors no longer even mention they are using it.
Jim Handy is a semiconductor and SSD analyst at Objective Analysis in Los Gatos, Calif.
