DeepSeek, the Chinese AI company, has just open-sourced the Fire-Flyer File System (3FS), a technology that helps AI models run faster while using fewer resources. The move comes as Western organizations have voiced doubts about the company's transparency. DeepSeek's “Open Source Week” event has drawn the attention of the international AI community, and many experts have praised the technology the company released.
DeepSeek Open-Sources 3FS, a File System That Accelerates AI
During “Open Source Week,” DeepSeek open-sourced 5 major software projects, the most notable of which is the Fire-Flyer File System (3FS). This technology optimizes AI processing, and DeepSeek uses it internally to train and deploy its AI models.
3FS is designed to take advantage of the performance of modern SSD storage and RDMA networking, optimizing AI workloads that need to process enormous amounts of data. The file system works without read caching and prioritizes random read requests, an important feature when GPUs constantly pull data from many different servers.
🚀 Day 5 of #OpenSourceWeek: 3FS, Thruster for All DeepSeek Data Access
Fire-Flyer File System (3FS) - a parallel file system that utilizes the full bandwidth of modern SSDs and RDMA networks.
⚡ 6.6 TiB/s aggregate read throughput in a 180-node cluster
⚡ 3.66 TiB/min…- DeepSeek (@deepseek_ai) February 28, 2025
According to Tom’s Hardware, 3FS can combine the bandwidth of thousands of SSDs and the network capacity of hundreds of storage nodes, which simplifies application code and makes use of standard storage APIs. On a 180-node server cluster, the system can reach read speeds of up to 6.6 TiB/s, while in the GraySort benchmark it hit 3.66 TiB/min on 25 nodes.
The AI company Perspective AI called these figures a “new milestone” and said 3FS could revolutionize big data processing, especially in AI, research, and data science.
DeepSeek Boosts Transparency as It Takes On Global Competition
Last year, DeepSeek announced Fire-Flyer 2, a high-performance AI architecture. Thanks to the combination of 3FS, HaiScale, and other software technologies, DeepSeek claims it reached 80% of the performance of Nvidia DGX-A100 servers at only 50% of the cost and with 40% less energy consumption.
The Fire-Flyer 2 system includes:
- 180 storage nodes, each with 16 SSDs of 16TB
- 2 NUCs at 200Gbps
- 10.000 Nvidia A100 GPUs connected over PCIe
DeepSeek organized “Open Source Week” to improve transparency after criticism from the West that the company lacks openness and credibility. DeepSeek has released the source code of several important software products, such as FlashMLA, DeepEP, and DeepGEMM, to spur community innovation and compete with the world's major AI companies.
This step not only lets DeepSeek demonstrate its capabilities but also makes the company one of the key players in the open-source AI market. By optimizing cost and resources in AI training, DeepSeek could become a formidable rival to large corporations such as OpenAI, Nvidia, and Google DeepMind.
