DeepSeek open-sources file system, claims it runs AI fashions quicker and extra effectively

Learn extra at:

What simply occurred? In response to Western organizations calling it “shady and untrustworthy,” DeepSeek launched “Open Supply Week.” Throughout final week’s occasion, the corporate launched a number of repositories to the open-source group, together with a extremely environment friendly file system. Many AI specialists reviewing the code have come away impressed.

Final week, DeepSeek launched 5 of its most superior software program repositories throughout its “Open Supply Week” occasion. The Chinese language AI agency unveiled a Linux-based file system it makes use of internally for AI coaching and inference workloads. The Fireplace-Flyer File System (3FS) boasts some spectacular efficiency benchmarks. Western AI corporations have taken notice and are exploring the repos. The corporate designed 3FS to speed up AI duties. The expertise leverages the options of recent solid-state storage models and RDMA networks, offering a shared storage layer to simplify the deployment of distributed purposes.

Tom’s Harware notes that DeepSeek’s 3FS code works with out read caching and prioritizes random learn requests since AI fashions operating on GPU nodes continually entry information snippets saved on servers. The file system can mix the throughput of 1000’s of SSD models and the community bandwidth of lots of of storage nodes, simplifying utility code and making use of normal storage API fashions.

The distributed file system can attain a 6.6 TiB/s combination learn throughput when utilized in a 180-node cluster, attaining a 3.66 TiB/min throughput on the GraySort benchmark (in a 25-node cluster). Startup firm Perspective AI praised DeepSeek’s figures as some “next-level” benchmarks, describing 3FS as a possible revolution for data-heavy workloads associated to AI, analysis, and extra.

In a paper printed final summer season, DeepSeek researchers described the options of the corporate’s customized Fireplace-Flyer 2 AI high-performance computing structure. Because of 3FS, HaiScale, and different parts of its software program stack, DeepSeek achieved 80 p.c of the efficiency of Nvidia’s DGX-A100 servers at 50 p.c of the value and utilizing 40 p.c much less vitality. Fireplace-Flyer 2 used 180 storage nodes with 16 16TB SSDs every, two 200Gbps NUCs, and 10,000 Nvidia A100 GPUs over PCIe.

DeepSeek created Open Supply Week to emphasise its transparency and community-based innovation after being criticized as shadowy and untrustworthy. The Chinese language firm is releasing many software program merchandise as open-source repositories, with key targets together with FlashMLA, DeepEP, DeepGEMM, and extra.


Turn leads into sales with free email marketing tools (en)

Leave a reply

Please enter your comment!
Please enter your name here