DeepSeek has released a new paper,??? ?? ???? 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]
Anaheim Honors Patti Hirahara with Proclamation
JANM Announces In
Mobile Showers Address Basic Need for Homeless
JACCC Celebrates 41st Anniversary with Kansha
Man Gets Life Sentence for Kidnapping, Torture, Mutilation
High Sloth NFTs Sold Out in 29 Minutes for $1.2 Million Dollars
Funds Being Raised to Repair Damage at Shofuso
KSCA Adds Tea Ceremony to May 28 Bazaar
Torrance Optometrist Receives ‘Best in South Bay’ Award for 9th Year
Salute for a 442nd Soldier
OBITUARY: Playwright Wakako Yamauchi, Remembered for ‘And the Soul Shall Dance’
SFVHBT Obon Festival Set for June 24
‘No No Girl’ Screening at ESGVJCC
2023 Japanese Cultural Bazaar at Buddhist Temple of San Diego
Challenger Astronaut Onizuka Remembered
More Than Names on a Wall
Honor on Parade
IT PAYS TO KNOW: Hiring In
OBITUARY: Mas Fujimoto, 92; Longtime Koyasan Troop 379 Scout Leader
Top Value and Huobi Launch a Blockchain Mining Fund
BinaryX Launches Strategy Game CyberChess With $500,000 Prize PoolVega Squadron qualify after beating LiquidMedia and Consultancy Startup PANONY Valued at $100MWest L.A. UMC Hosts Asian Cultural BazaarKulfi Finance Introduces Fixed Rate Lending and Borrowing Protocol on CardanoScreening of ‘The Motel’ at JANMBlockchain Economy Dubai Summit Is Just Around the CornerMintology Announces the Launch of New Brand Centric Claimable NFT PlatformSuper Gremlin announces Metaverse expansion by acquiring a prime location in DecentralandCopium Protocol NFT Pre Mint Goes Live The Year History Died EPA will now classify wood burning as carbon neutral, except it isn't Best smartwatch deal: Save $40 on the Fitbit Versa 4 New ExoMars photo shows Mars bathed in dramatic light Cows could be the largest mammals left on Earth in 300 years Astronomers are out looking for long Scott Pruitt proposes EPA limit agency's use of scientific studies No Re-Turning Point, U.S.A. Watch SpaceX send NASA's alien Best Fire TV Spring Sale deal: Save $150 on the Insignia 70
0.141s , 9827.1484375 kb
Copyright © 2025 Powered by 【??? ?? ????】Enter to watch online.DeepSeek reveals cost,Global Perspective Monitoring
