重磅！英伟达宣布 Rubin量产，液冷架构首次曝光

Core Viewpoint - NVIDIA has officially launched the Rubin platform, marking a significant technological advancement in the field of AI supercomputing, with full production announced by CEO Jensen Huang at CES 2026 [2]. Group 1: Rubin Platform Overview - The Rubin platform consists of six new chips that work together to create a powerful AI supercomputer, significantly reducing AI training time and lowering inference token generation costs [2]. - The Vera CPU, a core component of the platform, is built on a customized Armv9.2 architecture, featuring 88 cores and 176 threads, and supports NVIDIA's spatial multithreading technology [5]. - The Rubin GPU integrates a third-generation Transformer engine, providing 50 petaflops of NVFP4 computing power, with a single GPU bandwidth of 3.6 TB/s [8]. Group 2: Performance Metrics - The Rubin platform achieves 50 PFLOPS for inference and 35 PFLOPS for training, with HBM4 bandwidth reaching 22 TB/s, and NVLink bandwidth per GPU at 3.6 TB/s [6]. - The total number of transistors in the Rubin platform is 336 billion, representing a 1.6x increase compared to previous generations [6]. Group 3: Structural Innovations - The Rubin architecture features a complete redesign of the compute tray, transitioning to 100% liquid cooling and eliminating all cables, which enhances assembly efficiency and reduces the risk of operational errors [10]. - The cooling system utilizes a high-temperature water cooling technique, allowing for high power density heat dissipation while maintaining a water temperature of 45°C [12][15]. Group 4: Ecosystem and Partnerships - Major partners such as Lenovo and Supermicro are actively developing AI factories based on the Rubin platform, indicating a strong ecosystem support for the commercial deployment of advanced AI capabilities [14]. - The efficient liquid cooling architecture of the Rubin platform is expected to save approximately 6% of global data center electricity consumption [15]. Group 5: Comparison with Previous Platforms - The Rubin platform's power consumption is double that of the Grace Blackwell platform, yet it maintains the same cooling temperature, eliminating the need for expensive chiller systems in data centers [15]. - The GB300 platform, which was introduced in 2025, features a TDP exceeding 1.2 kW, showcasing the increasing energy demands of AI servers [16].