Data is being generated today at a rate far greater than ever imagined. In the past, people were the main source of data generation. There are now imaging devices, sensors, drones, connected cars, Internet of Things (IoT) devices, and pieces of industrial equipment that generate data in different ways and formats. However, we must not confuse data and information and it is vital to differentiate between the two terms. Currently, only a small part of the data collected is valuable enough to be considered a true asset. Let's take an imaging device as an example. In this case, what really matters is a minute of relevant activity, and not long hours of unnecessary video recording during which nothing important happens. As an analogy, "data" is the mine that people dig for the gold nugget, which is "information." The ability to turn this data into valuable information (the “dig,” to continue the analogy) can be called “analysis.”
The chart shown in Figure 1, produced by the analyst firm Statista, depicts the phenomenal increase in data storage capacity over the past decade. It predicts that by 2020 the demand for storage will exceed 42.000 exabytes. However, for the most part the stored data (estimates suggest at least 80%) is completely unstructured, making it difficult to use for analytical purposes.
Estimates indicate that only 5% of stored data is actually analyzed. If this unstructured data could be represented with metadata that describes it in the context of the analysis performed, a much larger amount of data could be analysed. This in turn significantly increases the value that organizations can generate from the data they hold. Artificial intelligence (AI) is a technology destined to have a powerful influence on modern society, and specifically on aspects such as e-commerce recommendations, natural language translations, financial technology, security, object identification/detection and even in the field of medicine, where potentially life-threatening cancer cells (as well as other abnormalities) can be quickly identified. Despite their diversity, all of these applications have a common thread in that we already have technology that can effectively sweep through huge amounts of unstructured data (videos, text, voice, images, etc.) and process it to obtain its true value. Specifically, we can use AI not only for the analytics process itself, but also for pre-processing of the raw unstructured data to provide the metadata with tags that can simply and accurately represent it.
This simplified database can be analyzed by analysis software at a higher layer and extract useful information. Organizations have been waiting for AI to make better use of the data they store, and until this stage they have remained “in the dark”. With that being said, we want to generate metadata to allow our analytics software to run more effectively and we have AI as a tool to create the metadata database from the huge unstructured database. Now we only have to transfer this enormous amount of data to our AI computing entities so that they can fulfill their task. Now, is this the correct way to do it? If we look at the two main points where data is generated and stored today, that is, in the “Cloud” and at the “Edge”, it becomes immediately clear that moving large amounts of data is very difficult. Expensive so should be avoided. In the Cloud, routing all of this data through the data center puts a heavy strain on the network infrastructure, consumes a lot of power, and increases latency levels (which adds to the overall processing time).
At the Edge, computing resources and available power are limited. The reduced network capacities of small devices located there will make it not feasible to upload large amounts of data to the Cloud for processing. In both cases, minimizing the amount of data being moved and relying on metadata is critical to optimizing operational efficiency. It will be much more effective if, instead of moving data, the mapping of the data can be done at the source, that is, where the data is located within the storage device itself. Solid State Drives (SSDs) already incorporate the fundamental elements necessary to function as computing entities. These are normally used for the operation of the drive, but can be reassigned to take over function-related tasks and to handle this labeling job, as well as to supplement the built-in hardware/software/firmware blocks that perform those functions.
One mode of operation may use drive free windows to perform scheduling tasks in the background. Another technique may be to process this data as it is being written to the drive. Consumption and cost savings – together with the minimization of data movement and a great reduction in latency, together with less network traffic – are some of the advantages that acceleration like this provides at the point of storage if it is used. apply in appropriate cases. The inherent scalability of this technique will allow enterprises and cloud service providers to expand the scope of their capabilities through the power of AI. At the Flash Memory Summit in Santa Clara in August, Marvell presented a revolutionary AI-based SSD controller concept that demonstrates how data tagging can be effectively executed without needing to access CPU processing resources on the host; in this way the problems of cost and latency described above are avoided.
Attendees were able to test the operation of Marvell's data center and customer's SSD driver ICs using the open source NVDLA (NVIDIA Deep Learning Accelerator) technology, how to leverage a proven AI model, compile it to the inference IP built into the AI and drill into a large database of unstructured data (eg, a video library) stored locally on the drive. From this you can generate tags and create a metadata database that appropriately represents the data in the context of the search. If the goal is to detect and recognize objects or scenes, the AI inference engine can explore video files stored on the drive and create metadata indicating when they appeared in the video. Thanks to this AI-enhanced storage technology, the metadata database can be stored locally on the SSD and made available to analysis software for necessary examination.
Consider, for example, a law enforcement agency looking for a suspicious “object” through endless hours of video files. They can load a ready-made model that knows exactly how to recognize an “object” and run inference for all available video content in parallel as a background task for all drives that store it. Any occurrence of this “object” would be flagged and tagged, making further analysis much easier. Similarly, consider how effective this architecture could be for something like background ChatBot analytics, where there is a large database of ChatBot calls that needs to be reviewed to improve quality of service.
It would be possible to assess when users were happy/unhappy with the responses received, or if the call was too long/too short. Once an AI model is created that knows how to follow these parameters, they could be picked up in the inference engines stored in the AI and the calls analyzed offline. Applications such as personalized placement of advertisements in video-on-demand services can also take advantage of its performance, such as searching for people or objects, and various use cases where intensive and I/O usage is used. take advantage of proximity to data.
Marvell's proposed AI SSD controller technology demonstrates how new data storage architectures can be implemented to handle the growing number of emerging "Big Data"-related applications that demand high computing power, without the need for expensive circuitry custom built integrations. By giving off-the-shelf SSD hardware access to auxiliary logic that dramatically increases its level of intelligence, vital tags and metadata can be taken directly for next-generation analytics tasks. There is no need to connect to a dedicated processing resource.
Implementing this alternative architecture to conventional centralized processing will make the entire procedure much more efficient. It requires only a minimum of available network bandwidth and prevents bottlenecks from forming. With AI accelerators built directly into inexpensive SSD controller ICs, analysis tasks can be completed quickly. It will also require less processing power and lower power consumption, as well as eliminating the need to build a complex ASIC from scratch. Using a programmable architecture also makes it much easier to update the AI models used, so their practical use can be addressed as they are developed.
