Serving technology enthusiasts for more than 25 years. TechSpot is a trusted source for tech advice and analysis.
What just happened? In reply to Western organizations calling DeepSeek “shady and untrustworthy,” DeepSeek has launched “Open Source Week.” Last week, the company released a number of repositories for the open-source community. This included a highly efficient filing system. Many AI experts who reviewed the code were impressed.
DeepSeek released its most advanced software repositories last week during its “Open Source Week” conference. The Chinese AI company unveiled a Linux file system that it uses internally to handle AI training and inference workloads. The Fire-Flyer File System 3FS boasts impressive performance benchmarks. Western AI companies are taking note and exploring the repos. The company designed 3FS accelerates AI tasks. The technology uses the features of solid-state storage and RDMA networks to provide a shared storage layer that simplifies the deployment of distributed apps.
Tom’s Harware notes DeepSeek’s 3FS works without Read caching prioritizes random read requests, since AI models running in GPU nodes constantly access data stored on servers. The file system can combine thousands of SSD units with the network bandwidth of hundreds storage nodes to simplify application code and apply standard storage API models.
6.6 TiB/s aggregate read throughput in a 180-node cluster
3,66 tib/min …– DeepSeek (@deepseek_ai) February 28, 2025
The distributed file system can reach a 6.6 TiB/s aggregate read throughput when used in a 180-node cluster, achieving a 3.66 TB/min throughput on the GraySor
Achieving a 6.6-TiB/s aggregate throughput with a 180 node cluster
Achieving a 3.66-TiB/min…
– DeepSeek @deepseek_ai February 28, 2025.
When used in a cluster of 180 nodes, the distributed file system is capable of achieving a 6.6-TiB/s throughput. Perspective AI, a startup company, praised DeepSeek for its “next-level” benchmarks and described 3FS as an important tool for data-heavy workloads in AI, research, etc. In a paper that was published last summer, DeepSeek researchers discussedfeatures of the company’s custom Fire-Flyer 2 AI architecture. DeepSeek’s software stack included 3FS, HaiScale and other components that allowed it to achieve 80 percent of Nvidia DGX A100 servers performance at a fraction of the cost and with 40 percent less power. Fire-Flyer 2, a storage node with 180 storage nodes, used two 200Gbps NUCs and 10,000 Nvidia GPUs A100 over PCIe.
DeepSeek launched Open Source Week in order to highlight its transparency and community based innovation, after being criticized for being shadowy and untrustworthy. The Chinese company releases many software products in open-source repositories. Key targets include FlashMLA, DeepEP and DeepGEMM.
