Skip to content

3FS

A distributed file system for AI workloads

Visit official site

What 3FS is

The system is described as a high-performance distributed file system for AI training and inference, combining SSD throughput with RDMA networking into a shared storage layer.

What you can do with it

  • Access storage in parallel across many SSDs and nodes
  • Rely on strong consistency through the CRAQ protocol
  • Use a standard file interface for AI workloads

Application scenarios

  • Store and quickly read training data for large AI training
  • Support LLM inference with KVCache
  • Manage intermediate data pipeline outputs during preparation

Pros & cons

✓ What we like

  • A standard file interface means no new storage API to learn
  • Large cluster deployment is supported

! What to watch out for

  • The design targets AI training and inference workloads
  • Data preparation, sample loading and checkpoints are named workloads

FAQ

What consistency protocol is used?

The material names the CRAQ protocol for strong consistency.

Do I need a new storage API?

The material states a standard file interface is provided.

Last reviewed: 2026-11-15

More AI coding tools tools

View all →

How we review