news

한어Русский языкEnglishFrançaisIndonesianSanskrit日本語DeutschPortuguêsΕλληνικάespañolItalianoSuomalainenLatina

[anthropic’s ceo also said earlier that the current ai model training cost is $1 billion, and in the next three years, this figure may rise to $10 billion or even $100 billion.]

the "10,000-card cluster" is regarded by the industry as the "entrance ticket" to this round of large-scale model competition. now, the "100,000-card cluster" has become a new high ground for technology giants to compete.

"more 100,000-card clusters will appear soon." on september 25, baidu group executive vice president shen dou said this at the baidu zhiyun conference.

he mentioned that in the past year, customers' demand for model training has surged, and the required cluster size has become larger and larger. at the same time, everyone's expectations for the continued decline in model inference costs have also increased. all of these have put forward higher requirements for the stability and effectiveness of gpu management. on the same day, baidu upgraded its ai heterogeneous computing platform baige 4.0, which has the ability to deploy and manage clusters of 100,000 cards.

in fact, the reason behind this round of generative ai explosion is partly due to the fact that “great effort brings miracles”. the industry has achieved a leap in the performance of large models by continuously increasing the computing power stack. therefore, the 10,000-card cluster is regarded by the industry as the “standard configuration” for entering the ai ​​core circle. but now, even 10,000 cards cannot fully meet the demand. not only baidu, but more and more industry giants are deploying 100,000-card clusters to pursue higher computing efficiency and large model performance.

at the recent yunqi conference, alibaba cloud demonstrated its new infrastructure for the ai ​​era, in which a single network cluster has expanded to the level of 100,000 cards. it is rebuilding an advanced ai infrastructure for the future from chips, servers, networks, storage to heat dissipation, power supply, data centers and other aspects.