DeepSeek open-sources file system, claims it runs AI models faster and more efficiently

Mar 4, 2025

DeepSeek open-sources 3FS for efficient AI tasks, enhancing AI model performance with 6.6 TiB/s throughput in a 180-node cluster.

DeepSeek open-sources file system, claims it runs AI models faster and more efficiently

: DeepSeek's Open Source Week introduced five advanced repositories, including the Fire-Flyer File System (3FS), designed to optimize AI training and inference workloads. The 3FS leverages SSD and RDMA technologies, reaching 6.6 TiB/s read throughput in a 180-node cluster. This innovation promises substantial performance improvements, matching 80% of Nvidia DGX-A100 servers at half the price. DeepSeek's transparency initiative responded to criticism, showcasing significant benchmarks.

During Open Source Week, DeepSeek released the Fire-Flyer File System (3FS), along with other advanced repositories, aiming to enhance AI model efficiency. 3FS is a Linux-based parallel file system that uses modern SSDs and RDMA networks to simplify deploying distributed applications and accelerate AI tasks.

This system achieves a notable 6.6 TiB/s aggregate read throughput in a 180-node cluster. On the GraySort benchmark, it performs at 3.66 TiB/min using a 25-node cluster, earning high praise from companies like Perspective AI for its potential to revolutionize data-heavy workloads.

Described in a paper by DeepSeek researchers, the Fire-Flyer 2 AI architecture reached 80% of the performance of Nvidia's DGX-A100 servers at a lower cost. This architecture incorporated 180 storage nodes with multi-terabyte SSDs and thousands of Nvidia A100 GPUs, showcasing DeepSeek's significant advancements in AI technology.