DeepSeek has released a new paper,eroticism 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]
Garden Creations
Did Trump's executive order just make everyone in the U.S. female?
NYT Connections hints and answers for January 18: Tips to solve 'Connections' #587.
NYT mini crossword answers for January 19, 2025
PHOTO JOURNAL: Senshin Smiles
Best robot vacuum deal: Save $170 on the Roborock Q5 Max+
TikTok ban: VPN not working for many in US as they attempt to bypass ban
Chicago Bulls vs. Portland Trail Blazers 2025 livestream: Watch NBA online
Araki’s Latest Film Premieres at Sundance
What is RedNote? The Chinese app, aka Xiaohongshu, TikTokkers are flocking to
Second Sunday Jam Session at GVJCI
NYT mini crossword answers for January 19, 2025
TRUMP and MELANIA crypto memecoins launch ahead of inauguration, chaos ensues
Atlanta Hawks vs. Boston Celtics 2025 livestream: Watch NBA online
Scent of Sulu
Lions vs. Commanders 2025 livestream: Watch NFL Playoffs
Lions vs. Commanders 2025 livestream: Watch NFL Playoffs
Best Uber deal: Uber Teen users can get free rides after failing a driver's test
‘Breaking Silences’: Intergenerational Reading at JANM
U.S. Supreme Court rules TikTok ban can move forward
Your thoughts as you watch your Instagram likes roll in67% of workers in the gig economy wouldn't join it againWe're sorry, but someone baked a cake with ketchup in itIt's exam results day in the UK and students are tweeting away their sorrowsJustin Bieber's Instagram is back. Thank goodness.U.S. government issues warning on McDonald's recalled wearable devicesU.S. government issues warning on McDonald's recalled wearable devicesEasyJet passengers liveRyan Lochte has now been dumped by four of his major sponsorsFake news reports from the Newseum are infinitely better than actual news Best Apple AirPods Pro 2 deal: Save $70 at Best Buy 25+ best Black Friday beauty tech deals to shop before Black Friday is over: Dyson, T3, Solawave Best Black Friday kitchen deals 2024: Save on appliances, cookware, more Sling TV Black Friday deal: Get 50% off and one month of AMC+ The best movies on Peacock for when you need some wild fun Miami Heat vs. Portland Trail Blazers 2025 livestream: Watch NBA online Best Black Friday deals on books at Walmart, Amazon Best Black Friday 2024 fitness tracker deals: Garmin, Fitbit, Oura Best Black Friday 2024 fitness tracker deals: Garmin, Fitbit, more Keys vs. Samsonova 2025 livestream: Watch Adelaide International for free
0.189s , 9991.34375 kb
Copyright © 2025 Powered by 【eroticism】DeepSeek reveals cost,Feature Flash
