Silicon Motion Technology Corporation (NasdaqGS: SIMO) has introduced the SM2524XT, a next-generation PCIe Gen5 DRAMless SSD controller designed for AI PCs and AI inference workloads.
The company said the new controller is optimized for KV Cache-intensive operations and on-device large language model (LLM) processing, where low latency and sustained random performance are becoming increasingly important for AI computing systems.
The SM2524XT features a four-processor-core architecture with PCIe Gen5 x4 connectivity and NAND interface speeds of up to 4,800 MT/s. According to Silicon Motion, the controller delivers sequential read speeds of up to 14 GB/s and random performance of up to 2.5 million input/output operations per second (IOPS).
Built on TSMC’s 6nm process technology, the controller is designed to improve both performance and power efficiency in AI PCs operating under thermal and power constraints.
Silicon Motion stated that the SM2524XT provides up to 25% higher performance per watt compared with its previous-generation controller platform. The company also said random performance improves by up to 25%, helping reduce latency and accelerate response times for fragmented data access patterns commonly associated with AI inference and KV Cache workloads.
As AI PCs evolve to support local AI agents and on-device generative AI applications, storage performance has become a growing bottleneck for inference responsiveness. Unlike conventional consumer SSD workloads, AI inference operations generate continuous streams of latency-sensitive random read and write requests that require stable throughput under sustained load conditions.
Nelson Duann said KV Cache has emerged as a critical factor in AI inference performance, increasing demand for sustained random throughput and low-latency storage access.
The SM2524XT integrates several proprietary technologies developed by Silicon Motion, including Separated Command Address (SCA) technology, advanced flash translation layer scheduling, and NANDXtend LDPC ECC technologies.
According to the company, these technologies improve parallel data processing efficiency, reduce latency interruptions, and maintain consistent performance during prolonged AI inference sessions.
