"The Vera Rubin platform, made up of six separate Nvidia chips, is expected to debut later this year, with the flagship server containing 72 of the company's graphics units and 36 of its new central processors. Huang showed how they can be strung together into pods with more than 1,000 Rubin chips and said they could improve the efficiency of generating what are known as tokens the fundamental unit of AI systems by 10 times."
"To get the new performance results, however, Huang said the Rubin chips use a proprietary kind of data that the company hopes the wider industry will adopt. This is how we were able to deliver such a gigantic step up in performance, even though we only have 1.6 times the number of transistors, Huang said. While Nvidia still dominates the market for training AI models, it faces far more competition from traditional rivals such as Advanced Micro Devices."
Nvidia's next-generation Rubin chips are in full production and claim up to five times the AI computing power of prior chips for chatbots and AI applications. The Vera Rubin platform comprises six separate chips, with a flagship server housing 72 graphics units and 36 central processors, and pods that can exceed 1,000 Rubin chips. Rubin configurations aim to improve token-generation efficiency by tenfold while relying on a proprietary data format and only 1.6 times more transistors. Increased competition from rivals and customers has prompted features such as context memory storage for longer conversations and upgraded networking with co-packaged connections.
Read at www.theguardian.com
Unable to calculate read time
Collection
[
|
...
]