英伟达官宣新一代GPU

Core Viewpoint - The article discusses NVIDIA's announcement of its new Rubin data center products at the CES, highlighting its advancements in AI hardware and the competitive landscape in the AI sector [4][5]. Group 1: NVIDIA's New Products - NVIDIA's CEO Jensen Huang announced the upcoming release of the new Rubin data center products, which are designed to accelerate AI development [4]. - The Rubin architecture consists of six independent chips, including the Rubin GPU and a new Vera CPU designed for "Agentic Reasoning," which is touted as the most advanced technology in the AI hardware field [5]. - The Rubin architecture reportedly offers a 3.5 times speed increase in AI model training tasks compared to the previous Blackwell architecture, with a fivefold performance improvement in running AI software [5]. Group 2: Performance and Cost Efficiency - The Rubin system can reduce the cost of generating inference tokens by up to 10 times and decrease the number of GPUs required for training mixture of experts (MoE) models by four times compared to Blackwell [5]. - The Vera CPU features 88 cores, providing double the performance of its predecessor, and is designed for high efficiency in large-scale AI operations [6]. - NVIDIA claims that the operational costs of systems based on the Rubin architecture will be lower than those based on Blackwell, as they require fewer components to achieve similar results [6]. Group 3: Market Position and Competition - NVIDIA aims to maintain its leading position in the AI accelerator market, despite increasing competition and concerns about the sustainability of AI spending growth [5][6]. - Major cloud providers, including Microsoft, are expected to be among the first to deploy the new hardware later in the year [6]. - The company remains optimistic about the long-term growth of the AI market, projecting it could reach trillions of dollars [6]. Group 4: Additional Hardware Features - The new hardware also includes networking and connectivity components, which will be part of the DGX SuperPod supercomputer and available as standalone products for modular use [7]. - This performance enhancement is necessary as AI models become more specialized, requiring the ability to process vast amounts of input and solve specific problems through multi-stage processes [7].