Since ChatGPT became popular, people have realized the huge commercial value brought by AI. AI smartphones are one aspect of it. AI smartphone users have requirements such as capacity, power efficiency, data efficiency, and security. However, the limited size, weight, and expected battery life of smartphones require lower power consumption and limit the capacity of storage systems. For this environment, Silicon Motion recently launched the SM2756 controller.
Instead of using a performance-optimized NVMe platform, the SM2756 uses the Universal Flash Storage (UFS) 4 specification optimized for mobile devices. This specification version gives the storage system a significant performance advantage over UFS 3.1. The SM2756 fully exploits this potential by providing a 2-channel HS-Gear-5 interface, adopting MPHY 5.0 technology and achieving a continuous read rate of up to 4.3 Gbytes/s. This means that a 3 billion parameter AI model can be loaded onto a smartphone in less than half a second.
To meet the capacity demands of AI smartphones, the SM2756 supports both Tri-Level and QLC 3D flash devices. It can manage up to 2 TB of storage. The use of TSMC 6 nm silicon and aggressive dynamic power management again helps with power efficiency. This controller will save nearly 60% of power when loading large parameter files compared to the power consumption of a similar UFS 3 controller.
Like the SM2508, the SM2756 uses extensive firmware algorithms to optimize data efficiency. While this is slightly different under UFS 4 than under NVMe, the impact on latency, actual transfer rates, and endurance is just as significant.
The SM2756 uses anti-hacking algorithmic code to address security concerns. This code prevents hackers from intervening during boot, ensuring data integrity and security on mobile devices.
In addition, the performance, capacity, power efficiency, and data efficiency of the storage subsystem all clearly impact the user experience of using the device. Meeting the storage subsystem requirements requires a large number of storage controller chips. The chip architecture, hardware implementation, and firmware must be optimized for these AI workloads. This requires detailed knowledge of NAND flash concepts and the operational details of individual flash chip families, acquired through decades of close collaboration with flash vendors. This is a project that system developers want to collaborate on with others.
Controller vendor independence is critical. In today’s highly dynamic global semiconductor market and uncertain supply chain environment, being locked into a single flash vendor due to storage controller limitations can be a very expensive mistake.
Silicon Motion’s decades of close relationships with all the major NAND flash vendors, a deep understanding of the management of data within storage arrays, and a clear understanding of the importance of data security are advantages that bring advantages to the new generation of edge AI controllers.