Core Viewpoint - Nvidia is expanding its business beyond GPUs, positioning itself as a comprehensive provider of AI infrastructure, including data center accelerators, hardware, and software solutions for physical AI applications like robotics and autonomous driving [3][4]. Group 1: Product Expansion - At the GTC conference, Nvidia introduced a variety of products, including data center accelerators, networking products, and open-source models, indicating a shift towards a broader AI infrastructure role [3]. - The introduction of the Groq 3 and Groq 3 LPX chips, which enhance the performance of Nvidia's Rubin platform, signifies a diversification in Nvidia's product offerings beyond traditional GPUs [7]. - The Groq 3 LPX chip can increase inference throughput by 35 times per megawatt when used with Rubin CPU and GPU, showcasing significant performance improvements [7]. Group 2: Chip Heterogeneity - Nvidia's strategy includes integrating LPU (Language Processing Unit) technology with GPUs to address the growing demand for faster inference speeds in large models, indicating a trend towards chip heterogeneity in AI workloads [10][11]. - Ian Buck emphasized that while GPUs will continue to dominate current AI applications, the combination of LPU and GPU will be crucial for next-generation AI workloads, particularly those involving trillion-parameter models [10]. - The industry is moving towards a heterogeneous computing environment, where different types of chips are needed for various workloads, as highlighted by AMD's collaboration with Meta to design semi-custom chips [10][11]. Group 3: Physical AI Development - Nvidia is making significant strides in physical AI, launching the Isaac simulation framework and the Cosmos model for robotics, which aims to unify synthetic world generation and physical AI reasoning [15][18]. - The company is focusing on open-source technologies to foster collaboration in the development of physical AI, as it believes that no single company can achieve this alone [15][19]. - Rev Lebaredian noted that the challenges in robotics are multifaceted, requiring advancements in hardware and software to make robots more functional and accessible [19].

要理解今天的英伟达可能比以往更不容易，但这家牵动着诸多AI领域发展的公司到底在如何勾勒AI的未来，仍值得探究。 英伟达业务扩充的信号变得明显。本届GTC大会上，英伟达发布的产品涵盖了数据中心加速器、机架、网络产品和多款开源模型。CUDA、GPU、LPU（语 言处理单元）、AI工厂、机器人、自动驾驶、开源模型等关键词在英伟达CEO黄仁勋的演讲中被频频提及。这家以GPU闻名的公司，如今将其定义为一家 包揽AI基础设施或AI工厂多个环节的厂商似乎更加合适。 即便只是在数据中心加速器这一环节，英伟达的产品类型也变得多样。Rubin平台在GPU之外，一款LPU也加入进来。原属于专用集成电路（ASIC）的LPU 与通用的GPU站在不同阵营，但英伟达拿下Groq的授权后，开启了两种芯片的联合。 而在以大型云厂商为客户的60%业务之外，看起来更为庞杂的40%业务中，英伟达也落下新子。物理AI中的自动驾驶和机器人成为两个重要抓手。为了部署 物理AI，英伟达不仅做硬件，还做自动驾驶平台和模型。 要理解今天的英伟达可能比以往更不容易，但这家牵动着诸多AI领域发展的公司到底在如何勾勒AI的未来，仍是值得探究的问题。GTC大会期间，第 ...