With the release of the DeepSeek-R1 LLM followed by image-generating Janus Pro, DeepSeek AI is on an open-source spree and they are not stopping as they announce Fire-Flyer File System (3FS) dedicated to AI-HPC operations.
3FS was released as part of the controversial AI company’s Open Source Week initiative with its key highlights being the astounding 7.3 TB/s aggregate read throughput within its own server data clusters. Surprisingly, this is not an entirely “new creation” as DeepSeek themselves has been using them since 2019.
A Linux-based parallel file system optimized for rapid random read speeds, 3FS ditches the usual traditional read caching to align specific needs of AI models where loads of training data were going in and out at random intervals consistently in a 1-time-only process.
Last year’s research paper released in August also detailed how 3FS benefited its Fire-Flyer 2 deep learning cluster by optimizing everything to achieve 80% of what NVIDIA’s DGX-A100 server solutions could but at only 50% cost and 60% power consumption, coming in at 6.6TBps with another 1.4TBps of read throughput for background training tasks while crushing a control set of Ceph file system at just 1.1TBps.
With it being live for public download and assessment and most importantly free, 3FS is expected to attract attention in the industrial segment despite it being a Chinese-developed software.
Source: Tom’s Hardware
