DeepSeek has released a new paper,Dear Utol: Week 1 Highlights Episode 7 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]
Best GoPro deal: Get the GoPro Max for under $400 at AmazonThe Hole by Nicolaia Rips115 Degrees, Las Vegas Strip by Meg BernhardUniversal Music is set to pull its songs from TikTokThe Paris Review Print Series: Shara Hughes by The Paris ReviewWhat the Review’s Staff Is Doing This Week: August 21–27 by The Paris ReviewPasolini on Caravaggio’s Artificial Light by Pier Paolo PasoliniPasolini on Caravaggio’s Artificial Light by Pier Paolo PasoliniSentences We Loved This Summer by The Paris ReviewAugust 7–13: What the Review’s Staff is Doing Next Week by The Paris ReviewNYT's The Mini crossword answers for January 31Wrong Turn by Natasha StaggScientists are now spotting whales from outer space using satellitesTrump’s political appointees will vet EPA’s scientific work before releaseWatch Jessica Laser Read “Kings” at the Paris Review Offices by The Paris ReviewFireworks: On Kenneth Anger and The Legend of Zelda by The Paris ReviewLifelines: On Santa Barbara by Jamie QuatroPasolini on Caravaggio’s Artificial Light by Pier Paolo PasoliniHow to watch 'Feud: Capote vs. the Swans': premiere date, streaming deals, and moreBest fitness deal: Get the Sunny Health & Fitness Mini Stepper for under $63 Porn site posts rape joke on Facebook, and people are enraged Everything you may have missed in the 'Stranger Things' Season 2 teaser At New York Fashion Week, supporting Planned Parenthood will be in style How GPS keeps up with the continent that's in constant motion The top 20 employee benefits and perks, according to Glassdoor Bold daughter steals the spotlight during Super Bowl post 'Wolfenstein 3D' mod asks you politely whether you want to shoot Nazis The evidence that Trump does, in fact, own a bathrobe '1984' is coming to Broadway to make you uncomfortable Fans relentlessly booed NFL commissioner Roger Goodell after the Super Bowl You can finally have a Cheeto that looks like Harambe ... for $99,999 Miley Cyrus announces on Instagram she prefers Hindu prayers to the Super Bowl UK mosques open up to visitors for food, tea, and a chance to talk South Korea to impose stricter phone battery safety regulations Bollywood film trailer starring a friendly ghost goes viral, gets 10 million views in a day For some reason, Justin Bieber is not over that PPAP song yet This space capsule looks so real that a NASA scientist reported it Woman at checkout reacts to girl's anxiety attack in the best possible way Trump becomes a mad medieval king in this brilliant Twitter parody Airbnb Super Bowl ad takes a dig at Trump with a beautiful message
2.499s , 10091.953125 kb
Copyright © 2025 Powered by 【Dear Utol: Week 1 Highlights Episode 7】,Warmth Information Network