With 2025 Computex Taipei focusing on the three major themes of "AI & Robotics", "Next-Gen Tech" and "Future Mobility", global technology giants have gathered to display their AI technology prowess, focusing on the core concept of "AI Next". The rapid deployment of AI applications has also accelerated urgent demand for high-efficiency storage technologies across various application scenarios. As the global leader of NAND Flash controllers, Silicon Motion plays a key role in AI ecosystem development.
Meeting Diverse Storage Requirements, from Low Latency and Power Efficiency to High Data Throughput, to Support Edge AI Growth
"The emergence of DeepSeek has greatly lowered the threshold for AI applications," pointed out Mr. Kou, President and CEO of Silicon Motion. As an open source technology, DeepSeek has been able to reduce the cost of language model training. It has gradually subverted the industry's traditional views on AI and led to the accelerating popularization of edge applications. He emphasizes that a wave of AI adoption has already begun for devices from smartphones and laptops to wearable devices, and that storage technologies are crucial in supporting this revolution.
In his analysis of AI storage architecture, Mr. Kou remarked that the storage system requirements for each stage of the implementation process differ when implementing AI applications in various scenarios, from initial data ingestion to the preparation, training, and inference stages. For instance, data ingestion requires import of a large amount of data, meaning that high write throughput is required. On the other hand, low latency performance and support for a wide variety of I/O sizes has greater importance in the model training stage. Although these requirements vary, the overall architecture must still possess five core characteristics: high data throughput, low latency, low power, scalability, and high reliability, in order to meet the needs of AI applications.
In response to the massive data demands of AI applications, Silicon Motion leads innovation in storage technologies by upgrading NAND controller technology. Mr. Kou said that data application processes can be effectively optimized through hierarchical management and smart identification mechanisms. Flexible data placement (FDP) technology can also serve to improve efficiency and durability, while also offering the advantages of being low latency and low cost. For data security and reliability, the product also adopts advanced encryption standards and a tamper-proof hardware design. In combination with end-to-end data path protection mechanisms and Silicon Motion's proprietary NANDXtend™ technology, this enhances data integrity and prolongs the SSD's lifespan. In addition, Silicon Motion supports 2Tb QLC NAND and 6/8-Plane NAND, combining smart power management controllers (PMC) with advanced process technology to effectively reduce energy consumption while improving storage density.
Not only that, it can also be paired with Silicon Motion's unique PerformaShape technology, which utilizes a multi-stage architecture algorithm to help optimize SSD performance based on user-defined QoS sets. Together, FDP and PerformaShape can not only help users effectively manage data and reduce latency, but also significantly improve overall performance by approximately 20-30%. These technologies are specifically suited for AI data pipelines in multi-tenant environments, including key stages such as the data ingestion, data preparation, model training, and inference processes.
Creating Comprehensive Solutions to Realize Customer AI Applications Across Cloud and Edge Computing
In response to data center and cloud storage needs, Silicon Motion has launched the world's first 128TB QLC PCIe Gen5 enterprise SSD reference design kit. By adopting the MonTitan SSD development platform, which comes equipped with an SM8366 controller, it is able to support PCIe Gen5 x4, NVMe 2.0, and OCP 2.5 standards. With a continuous read speed of over 14 Gb/s and a random access performance of over 3.3 million IOPS, it boasts a performance improvement of over 25%. This design is able to speed up training of large language models (LLM) and graph neural networks (GNN) while also reducing AI GPU energy consumption, allowing it to meet high-speed data processing demands.
For edge storage solutions, Mr. Kou stated that the number of edge devices with AI capabilities will grow rapidly. He forecast: "The AI humanoid robot market will see explosive growth in the next 5 to 10 years." Systems at different levels have different storage requirements. For example, at the sensor level, data needs to be processed and filtered in real time to ensure accurate data sensing, while decision-making relies on multi-modal fusion reasoning, which entails more demanding storage performance and data integration capabilities. Meanwhile, at the execution level, various calibration parameters must be stored to enable the robot to act and think more similarly to humans. In response, Silicon Motion has actively deployed NVMe SSD, UFS, eMMC, and BGA SSD storage solutions, and values greater cross-industry collaboration to build a shared eco-system, in order to promote the further evolution of smart terminal storage technologies.
Additionally, Silicon Motion has launched a variety of high-efficiency, low-power controller to meet the AI application needs of edge devices: The SM2508 PCIe Gen5 controller is designed for AI laptops and gaming consoles, featuring up to 50% lower power consumption compared to similar products. The SM2324 supports USB 4.0 high-speed portable storage devices up to 16TB in size. The SM2756 UFS 4.1 controller has a 65% higher power efficiency compared to UFS 3.1, providing an excellent storage experience for AI smartphones. In response to the urgent need for high-speed and high-capacity storage required for self-driving cars, Silicon Motion has also joined hands with global NAND manufacturers and module makers to jointly create storage solutions for smart automobiles.
"Storage technology undoubtedly acts as a core link in the AI ecosystem," emphasized Mr. Kou. Taiwan has a complete and highly integrated semiconductor and information and communications industry chain. It is capable not only of building AI servers, but also possesses great potential for promoting the development of AI applications. He believes that more practical AI edge computing devices and groundbreaking applications will be launched at a rapid pace in the future, and that storage solutions will face increasingly demanding requirements due to challenges in processing massive amounts of data. Silicon Motion will continue to use technological innovation as a driving force to actively support AI development.
Mr. Kou expressed that the fast-paced development of generative AI has led to lower barriers to adoption for related applications. Silicon Motion aims to satisfy the market's needs through offering a diverse range of high-efficiency, low-power storage solutions.
Photo: Silicon Motion Technology
Article edited by Jerry Chen
