DeepSeek has released a new paper,why sex video with co-founder Liang Wenfeng credited as a contributor, detailing how its latest large language model DeepSeek-V3 achieves efficient training and inference using only 2,048 H800 GPUs – significantly fewer than the tens of thousands typically required. The team attributes this efficiency to four key innovations: memory optimization through multi-head latent attention (MLA), computational savings via a Mixture-of-Experts (MoE) design with FP8 precision, communication improvements using a multi-plane network topology, and faster inference through multi-token prediction (MTP). With MLA, KV cache memory usage is cut to just 70KB per token, up to 1/7 that of competing models. MoE architecture activates only 37 billion of the model’s 671 billion parameters per forward pass, reducing training costs by 90% compared to dense models. FP8 training further halves compute and memory usage, with minimal accuracy tradeoff. Beyond the model, the paper also outlines five future directions for AI hardware design, advocating for tighter integration between software and hardware to address memory, compute, and networking bottlenecks. [36Kr, in Chinese]
Related Articles
2025-06-26 19:56
2740 views
NYT Connections hints and answers for May 10: Tips to solve 'Connections' #699.
Connectionsis the one of the most popular New York Times word games that's captured the public's att
Read More
2025-06-26 19:52
634 views
J.K. Rowling has apologised again for killing someone you loved in 'Harry Potter'
You have a lot to answer for, J.K.SEE ALSO: J.K. Rowling has spoken, owls are bet
Read More
2025-06-26 18:20
1767 views
Drop everything, Gru's 'gorls' meme is the funniest thing on Twitter
You may have seen the Gru's plan meme making its rounds, perfectly describing when your plans just g
Read More