Deepseek recently organized the Open Source Week event to address criticisms from Western organizations regarding their reputation for being “shady and unreliable.” During this event, the company unveiled five of its advanced software storage solutions, prominently featuring the Fire-Flyer File System (3FS), which is integral to its AI training and inference processes.
The 3FS file system is specifically designed to optimize AI workloads, garnering significant attention from industry experts. It leverages the capabilities of modern solid-state drives (SSDs) and Remote Direct Memory Access (RDMA) networks to create a shared storage layer, which simplifies the deployment of distributed applications. Experts from Tom’s Hardware have commended 3FS for its performance, noting its ability to operate without a reading cushion while prioritizing random reading requirements.
This advancement is particularly crucial for AI models that frequently retrieve data from servers. The distributed file system boasts impressive performance metrics, achieving a combined reading speed of up to 6.6 tib/sec when utilizing 180 buttons, and reaching 3.66 tib/minute on the Graysort standard within a 25-button cluster.
The startup company Perspective has recognized these performance figures as indicative of a “next level” advancement, praising 3FS as a potentially transformative technology for managing large data volumes associated with AI research.
In a previous article published last summer, researchers at Deepseek highlighted the high-performance capabilities of their Fire-Flyer 2 AI system. Through the use of 3FS and other components within their software stack, Deepseek has achieved 80% of the performance of NVIDIA’s DGX-A100 server, while incurring only 50% of the cost and consuming less than 40% of the energy.
With Open Source Week, Deepseek aims to promote transparency and foster innovation within the community by releasing various software products in an open-source format, including initiatives like Flashmla, Deepep, and Deepgemm.