Deepseek, a Chinese -based artificial intelligence company, presented five different software reposles to the open source community last week. This move draws attention by Western organizations as an effort to emphasize transparency by Deepseek, who faces criticisms such as “questionable” and “unreliability”. The Fire-Flyer File System (3FS), which is designed for AI training and inferences, attracted great attention among the software announced.

Deepseek surprised once again

3FS is a Linux -based file system developed by Deepseek for AI workloads in internal use. Using the characteristics of modern solid state drives (SSD) and RDMA networks, this system simplifies the distribution of distributed applications. The 3FS, which attracts attention with its high performance standards, has been optimized to meet the continuous data access needs of AI models working in GPUs. However, Deepseek’s design gives priority to random reading demands and does not use reading cache.

The company’s file system can reach a total reading band width of 6.6 TIB/S when used in a cluster of 180 nodes. In the Graysort Benchmark, he attracted attention by reaching a 3,66 TIB/minute reading speed with a 25 -knoted cluster. Startup company Perspective AI describes Deepseek’s performance figures as “next level” and says that 3FS offers a revolutionary solution for AI and data -density workloads.

Deepseek announced the features of the high-performance AI architecture called Fire-Flyer 2 last summer. Thanks to the other elements of the 3FS, Haiscale and software pile, 80 percent of the performance of Nvidia’s DGX-A100 servers using 50 percent and 40 percent less energy. The Fire-Flyer 2 used 180 storage nodes with a 16TB SSD, two 200GBPP Nuc and 10,000 Nvidia A100 GPUs via PCIE.