Serving tech followers for over 25 years.
TechSpot methodology tech diagnosis and advice you can trust.
What correct came about? Essentially based on Western organizations calling it “shady and untrustworthy,” DeepSeek launched “Open Source Week.” All over closing week’s match, the corporate launched several repositories to the originate-source neighborhood, alongside side a highly efficient file system. Many AI specialists reviewing the code have contrivance away impressed.
Final week, DeepSeek launched five of its most developed instrument repositories all over its “Open Source Week” match. The Chinese AI agency unveiled a Linux-primarily primarily based file system it uses internally for AI practicing and inference workloads. The Fireplace-Flyer File System (3FS) boasts some spectacular performance benchmarks. Western AI companies have taken point to and are exploring the repos. The company designed 3FS to tempo up AI tasks. The skills leverages the parts of current proper-relate storage devices and RDMA networks, providing a shared storage layer to simplify the deployment of disbursed functions.
Tom’s Harware notes that DeepSeek’s 3FS code works with out be taught caching and prioritizes random be taught requests since AI objects working on GPU nodes constantly fetch admission to data snippets kept on servers. The file system can mix the throughput of hundreds of SSD devices and the community bandwidth of a total bunch of storage nodes, simplifying utility code and applying long-established storage API objects.
🚀 Day 5 of #OpenSourceWeek: 3FS, Thruster for All DeepSeek Recordsdata Win admission to
Fireplace-Flyer File System (3FS) – a parallel file system that utilizes the stout bandwidth of current SSDs and RDMA networks.
⚡ 6.6 TiB/s combination be taught throughput in a 180-node cluster
⚡ 3,66 tib/min …– Deepseek (@deepsek_ai)”https://twitter.com/deepseek_ai/status/1895279409185390655?ref_src=twsrc%5Etfw”>February 28, 2025
The disbursed file system can reach a 6.6 TiB/s combination be taught throughput when archaic in a 180-node cluster, achieving a 3.66 TiB/min throughput on the GraySort benchmark (in a 25-node cluster). Startup company Standpoint AI praised DeepSeek’s figures as some “next-level” benchmarks, describing 3FS as a potential revolution for data-heavy workloads connected to AI, research, and more.
In a paper revealed closing summer time, DeepSeek researchers described the parts of the corporate’s custom Fireplace-Flyer 2 AI excessive-performance computing structure. Attributable to 3FS, HaiScale, and other facets of its instrument stack, DeepSeek achieved 80 percent of the performance of Nvidia’s DGX-A100 servers at 50 percent of the price and the exercise of 40 percent much less energy. Fireplace-Flyer 2 archaic 180 storage nodes with 16 16TB SSDs each and every, two 200Gbps NUCs, and 10,000 Nvidia A100 GPUs over PCIe.
DeepSeek created Delivery Offer Week to emphasise its transparency and neighborhood-primarily primarily based innovation after being criticized as dusky and untrustworthy. The Chinese company is releasing many instrument merchandise as originate-source repositories, with key targets alongside side FlashMLA, DeepEP, DeepGEMM, and more.
