At this event, DeepSeek announced five of its advanced software repositories, most notably the Fire-Flyer File System (3FS), which the company uses for AI training and inference workloads.
DeepSeek has created many breakthroughs in AI technology.
Designed to optimize AI tasks, the 3FS file system has garnered attention from many experts in the field. This technology leverages the features of modern solid-state storage units and RDMA networks, creating a shared storage layer that simplifies the deployment of distributed applications. Experts at Tom's Hardware appreciate that 3FS code operates without read caching and prioritizes random read requests.
3FS is the revolution that DeepSeek has brought about.
According to Tom's Hardware , this is crucial for AI models that frequently access data from servers. As a result, this distributed file system is capable of achieving aggregate read throughput of up to 6.6 TiB/second when operating in a cluster of 180 nodes and 3.66 TiB/minute on the GraySort benchmark in a cluster of 25 nodes.
The startup Perspective AI also hailed these metrics as "the next level," describing 3FS as a potential revolution for data-intensive workloads related to AI and research.
What is DeepSeek, and who is behind it?
In a paper published last summer, DeepSeek researchers presented the features of their custom Fire-Flyer 2 AI high-performance computing architecture. Thanks to 3FS and other elements in the software stack, DeepSeek achieved 80% of the performance of Nvidia's DGX-A100 server at only 50% of the cost and with 40% less power consumption.
Through Open Source Week, DeepSeek aims to highlight transparency and community-driven innovation, while releasing numerous software products as open-source repositories, including projects such as FlashMLA, DeepEP, and DeepGEMM.
Source: https://thanhnien.vn/deepseek-cong-bo-dot-pha-ve-cong-nghe-ai-185250304091501089.htm
Comment (0)