Nvidia(US:NVDA) 虎嗅APP·2025-03-19 10:18Core Viewpoint - NVIDIA's recent GTC conference showcased the launch of its new generation of AI chips, emphasizing the importance of inference efficiency over sheer computational power, with DeepSeek emerging as a significant player in this landscape [2][4][6]. Group 1: AI Chip Developments - NVIDIA introduced the Blackwell Ultra chip, which is set to deliver 20 petaflops of AI performance and features 288GB of HBM3e memory, marking a significant upgrade from its predecessor [10][12]. - The Blackwell Ultra chip will support various AI tasks, including pre-training, post-training, and inference, making it a versatile platform [10][12]. - The upcoming Rubin chip, expected in late 2026, will offer performance improvements of up to 900 times compared to the Hopper architecture, with capabilities reaching 3.6 ExaFLOPS for inference tasks [19][20][23]. Group 2: Inference Efficiency - The conference highlighted that the future of AI competition will hinge on achieving the lowest inference costs and highest efficiency, rather than merely increasing model size [6][4]. - NVIDIA's DeepSeek-R1 model achieved a throughput of over 30,000 tokens per second, showcasing a 36-fold increase in throughput since January [49][50]. - The company aims to optimize its entire inference ecosystem, integrating advanced tools to enhance performance across various frameworks [50][58]. Group 3: Networking Infrastructure - NVIDIA introduced the Spectrum-X and Quantum-X silicon photonic switches to enhance AI factory connectivity, significantly reducing energy consumption and operational costs [30][32]. - The new networking technology is designed to support millions of GPUs across sites, addressing the growing demand for bandwidth and low latency in AI applications [29][34]. Group 4: AI Factory Concept - NVIDIA's vision for the future includes the concept of AI factories, where every industry will operate both a physical factory and an AI factory, with Dynamo serving as the operating system for these AI environments [36][35]. - The company is positioning itself as a leader in transforming GPU computing into a foundational infrastructure for various industries, moving beyond traditional chip manufacturing [54][58]. Group 5: Robotics and AI Integration - The conference featured the introduction of the Isaac GR00T N1, a humanoid robot model that utilizes advanced AI frameworks for real-world applications [41][43]. - NVIDIA's collaboration with Google DeepMind and Disney Research on the open-source physics engine Newton aims to enhance robotic capabilities and AI learning [45][46].