On 20 November 2024, the Ballroom on the second floor of Raffles Perennial Shenzhen held the "Intelligent Computing Leads the Storage Core"-- MTS 2025 Storage Industry Trends Seminar. The symposium invited leading figures in the field of storage and AI to discuss the changes and challenges of storage technology in the AI era. As the guest speaker, Mr. Xiting Duan, Senior Vice President of CAS Business Group of Silicon Motion Technology, shared the keynote speech of "AI Data Efficiency, Key Storage Technologies", and deeply analyzed the four core capabilities faced by storage technology in the AI multimodal era. China Exportsemi summarizes and sorts out this part of the content sharing for you.
1. The relationship between AI and storage in the multimodal era
Mr. Duan Xiting opened the video to show the close connection between AI and storage technology, and emphasized the fundamental position of storage in the field of AI. He pointed out that although storage technology is relatively less discussed in the AI industry, it plays an integral role in the rapid development of AI. Whether it is the acquisition of big data or the training and inference of AI models, storage technology plays a vital role in the whole process.
With the deepening of AI technology, AI is no longer limited to a single data modality, but has gradually entered the stage of multimodality. Unimodal AI can only process one type of data (such as text or images), while multimodal AI is capable of processing and analyzing different types of data (such as text, speech, images, videos, etc.) at the same time. Duan Xiting explained that this shift has placed unprecedented demands on storage technology, which must not only support the flow of massive amounts of data, but also be able to handle complex data in multiple formats.
2. Four core capabilities: AI and storage: challenges and solutions
In his speech, Duan Xiting analyzed the four core capabilities of storage technology in the AI multimodal era, which will directly affect the efficiency and development of AI applications.
Figure: Xiting Duan, Senior Vice President of CAS Business Group of Silicon Motion Technology, believes that storage has four key roles in the AI ecosystem
1. High-capacity storage: Dealing with the challenges of data explosions
With the popularity of AI applications, the amount of data has exploded. Duan Xiting pointed out that with the amount of data per person per day reaching 10GB in 2024, this data volume is expected to increase to 100GB by 2034, and the total amount of global data is expected to exceed 200 zettabytes. To cope with this growth, the capacity requirements for storage devices will continue to expand. Applications in the fields of AI deep learning, image recognition, and autonomous driving require a large amount of data storage support, and the storage technology must have sufficient capacity to meet the massive data processing of AI applications.
In this context, the emergence of new storage technologies such as QLC NAND (Level 4 Cell Flash) provides a viable solution for high-capacity storage. By increasing storage density and reducing costs, QLC NAND is able to improve storage efficiency while meeting the demand for high-capacity storage.
2. Data security: Ensuring privacy and data integrity
In the era of AI multimodality, data security has become more and more important. Duan Xiting pointed out that with the popularity of edge computing, data is not only limited to cloud storage, but also needs to be transmitted and shared between various devices. Data in AI applications often contains sensitive information, and how to protect data privacy and ensure data integrity will become a core challenge for storage technology.
Especially in AI edge computing, data may involve smart terminals, IoT devices, etc., and how to ensure the security of data in the process of storage and transmission has become a major problem in the development of storage technology. Duan Xiting mentioned that in the application of AI, storage must not only provide efficient data access capabilities, but also have strong encryption and anti-leakage functions to ensure data security.
3. Data efficiency: Improve AI computing performance
Duan Xiting further elaborated on the key role of data efficiency in AI applications. In the training process of AI models, storage technology must have the characteristics of high throughput and low latency, and can provide efficient data flow in the process of data input and output. In particular, advanced AI applications such as large language models (LLMs) require storage systems that can not only quickly process massive amounts of data, but also maintain efficient data access.
For the AI field, storage efficiency is closely related to data processing capabilities. Duan Xiting said that the efficiency of reading and writing data in AI training directly affects the speed and accuracy of the entire computing process. Therefore, future storage technologies will pay more attention to the high speed of data transmission, especially in computing platforms such as GPUs and NPUs, and the performance of storage systems will directly determine the speed of AI computing.
4. Multi-modal storage: meet the needs of complex data formats
With the continuous development of AI technology, a single data format can no longer meet the needs of applications. Duan Xiting pointed out that AI technology is moving towards a multimodal era, and storage technology must be able to support different types of multimodal data. Different from traditional data storage, AI multimodal storage requires the system to be able to process data in multiple formats such as images, videos, voices, and texts at the same time.
For example, in the field of autonomous driving, vehicles need to process data from different sensors (such as lidar, camera, radar, etc.) in real time, and fuse and analyze these multimodal data. Storage technology needs to provide efficient data processing capabilities to ensure that different types of data can be quickly stored, retrieved, and analyzed. Duan Xiting emphasized that storage technology should not only cope with the diversity of data formats, but also ensure storage efficiency and high performance of the system.
1. AI edge computing: new requirements for storage technology
In addition to data centers, the rapid development of AI edge computing has also brought new challenges and opportunities to storage technology. Duan Xiting pointed out that with the popularization of 5G technology, AI will gradually expand to the edge, and storage devices need to achieve efficient and fast data processing on terminal devices. In edge computing scenarios, due to space, power consumption, and cost constraints, AI storage devices must be lightweight and efficient to ensure that they can support the computing needs of edge devices.
In edge computing, storage devices not only need to provide local storage, but also need to collaborate with the cloud to complete distributed storage and computing of data. This means that the storage technology needs to have the ability to have low latency and high throughput, and it also needs to support a distributed architecture to ensure efficient and smooth data transfer and processing between the cloud and the edge.
2. Summary: The development trend of storage technology in the AI multimodal era
Through this speech, Duan Xiting deeply discussed the four core capabilities faced by storage technology in the AI multimodal era, and analyzed how storage technology continues to innovate and optimize in terms of large capacity, data security, efficiency, and multimodal storage. In the future, with the continuous development of AI technology, storage technology will continue to provide strong support for the application of AI and become a key force to promote industry innovation and development.
Silicon Wing has played an important role in the continuous progress of storage technology, especially in the field of large-capacity, high-performance and high-security storage solutions, which has become an important driving force for the development of the industry. With the advent of the AI multimodal era, the storage industry will face greater challenges and opportunities, and the progress of storage technology will directly affect the widespread development and popularization of AI applications.
At the end of the speech, Duan Xiting introduced the comprehensive strategic layout of Silicon Motion Technology in the field of AI storage, which mainly focuses on four major fields: data centers, AI smartphones, AI notebooks and smart cars, and cooperates with the world's top CPU and GPU manufacturers to meet the diverse needs of the AI era. In the data center, Huirong launched the MonTitan SM8366 SSD main controller, which is designed for AI, with a 20% performance improvement, using a PCIe Gen5 interface, which has been verified by large language models. In terms of AI smartphones, the UFS 4.1 SM2756 main controller improves power consumption efficiency, increases battery life by 1.5 times, improves performance, and supports more stable AI applications. In the field of AI notebooks, the SM2508 PCIe Gen5 SSD main control data efficiency is increased by 3 times, the power consumption is reduced by 50%, and the performance is increased by 70%, and mass production is expected in December. In the field of smart cars, the PCIe Gen4 SSD master can handle multiple data streams, reduce power consumption by 30%, and is ASPICE and ISO certified to ensure high reliability and performance. Silicon Wing is committed to becoming a one-stop service provider in the field of AI storage, providing comprehensive solutions for global and local partners.
Figure: Silicon Motion Technology's comprehensive strategy for AI storage
All in all, the success of this seminar not only gave the participants an in-depth understanding of the future trend of storage technology, but also provided valuable ideas and inspiration for practitioners in the storage industry. The integration of AI and storage technology will surely lead the industry into a new stage of development.
