File-Semantic SSD: a file system built right into the drive—and why this isn’t news to gamers

Headlines claim that the new SSDs “don’t need a processor” and that they’ll speed up gaming PCs as well. The actual announcement is more specific and technically interesting: FADU, a South Korean SSD controller designer, has unveiled a drive that manages the file system on its own, allowing graphics cards and network cards to access files without the server’s processor getting in the way. This is an architectural demonstration for data centers with artificial intelligence, not a drive that is available for purchase.

What FADU Presented

FADU unveiled a device called the File-Semantic SSD (FS-SSD) at the Open Compute Project Global Summit in San José in mid-October 2026. It is built on the company’s sixth-generation PCIe controller. Traditionally, SSDs store data in blocks, and the file system—which organizes these blocks into files, directories, and permissions—runs in the operating system on the processor. The FS-SSD performs these tasks—namely, file lookup, permission checking, and translating file paths to addresses in flash memory—within its own firmware.

How to Access the Drive

According to FADU’s descriptions, it has encapsulated NFS version 3 network file system operations into standard NVMe commands. The graphics card can prepare file requests in its own memory, and the network card can insert them directly into the disk queue. Data then travels between the disk and the card directly via PCIe. The disk also supports network transfers via NVMe over Fabrics.

One administrator instead of two

The company combines a file system with a layer that manages flash memory. It is derived from the well-known F2FS file system. This integration is intended to eliminate “double erasure,” in which erasure at the file system level and erasure at the disk level reinforce each other.

Why is this done?

FADU reports that the file system and block layers consume approximately 30% of processor cycles during random reads for graphics cards (GPUDirect Storage). For large servers running AI workloads, this is a significant overhead that the company aims to eliminate. The long-term goal is serverless storage: a set of such drives connected to a PCIe switch, managed by a DPU network card, without a server processor.

What the metrics show

In its preliminary test, the company reported that when accessing files remotely via NFS and NVMe, performance remained at approximately 460,000 operations per second, even though the server reduced the number of cores from 16 to one. The test was conducted on a drive with a sixth-generation FC6161 controller in a host computer with PCIe 5.0, and these are the company’s own preliminary measurements. One analysis raises concerns about the methodology and recommends monitoring the topic rather than basing purchasing decisions on it.

When might this appear?

• Results on a host computer with PCIe 6.0 are expected by the first quarter of 2027.
• Samples should be available by the third quarter of 2027.
• Standardization has not yet begun. FADU plans to open a discussion on a command set for “file-based” semantics within NVM Express and OCP Storage.

What is not confirmed

• “An SSD doesn’t need a processor.” The drive has its own controller and firmware. The point is that the host processor doesn’t have to interfere with the data path. Overall, the computation is shifted to the drive.
• Speeding up gaming PCs. Sources describe data centers and AI workloads. We haven’t found any information on the impact on home computers, so this is speculation.
• Product availability. This is an architectural demonstration, and the technology is not intended for general purchase.
• Independent tests. All performance data comes from the company.

Security and Open Questions

When the file system runs on a disk, permissions must be addressed. According to the description, the company plans to use capability-based authentication because user identifiers are part of the command and a client could forge them. Only independent tests will show how well this will work in practice. It also remains to be seen whether the new command set will become a standard or remain a solution specific to a single vendor.

Conclusion

File-Semantic SSD is an interesting idea for reducing the CPU load when graphics cards are saturated with data: instead of the CPU, the disk handles the files. So far, it’s a proof of concept for data centers, with samples planned for 2027. For the average user, it’s primarily an example of the direction in which data storage might evolve—not a technology for home computers.