South Korean controller designer FADU has introduced a new approach to enterprise flash storage aimed at removing host CPU bottlenecks in artificial intelligence workloads. Unveiled at the OCP Global Summit, the File-Semantic SSD (FS-SSD) moves file lookup, access permissions, and file-to-address translation directly into firmware running on the company's sixth-generation controller. Rather than treating storage purely in traditional blocks, the drive is engineered to understand files natively, allowing GPUs, DPUs, and network cards to access data without requiring a host file system or storage-server CPU in the data path.
FADU's design integrates network file system requests, specifically NFSv3, directly inside standard NVMe commands. This allows commands to originate either locally within the same server or across the data center via NVMe over Fabrics. By utilizing a stateless protocol, the drive avoids tracking which client has specific files open. The architecture combines file mapping and flash translation into a single layer based on F2FS, a Linux file system optimized for flash memory. According to FADU, this consolidation reduces write amplification by 63 percent compared to running F2FS on a standard host, though these specific figures rely on simulated tests using NVMeVirt software.
While the architectural pitch seeks to eliminate the roughly 30 percent of CPU cycles typically consumed by host file systems during random-read workloads, practical deployment remains on the horizon. Preliminary tests conducted on a PCIe 5.0 host demonstrated steady remote file throughput of approximately 460,000 operations per second even when scaling down server resources. However, physical hardware samples featuring this advanced firmware and controller design are not expected to reach partners until the third quarter of 2027.
Comments