Alphabet(US:GOOG) 中金点睛·2025-12-08 23:37Core Viewpoint - The launch of Google TPUv7 signifies a shift towards self-developed ASIC clusters, enhancing hardware value through heterogeneity and restructuring, which is expected to accelerate the growth of the AI computing hardware market, including PCB, liquid cooling, and power supply components, with projected market sizes reaching $21.65 billion, $20.18 billion, and $18.39 billion by 2027 respectively [2][4]. Group 1: TPU Architecture Evolution - Google has evolved its TPU architecture over the past decade, transitioning from TPU v1, a pure inference co-processor, to TPU v7, which features significant advancements such as dual-chiplet packaging and enhanced linear acceleration in large-scale clusters [3][10]. - The TPU v7 architecture includes 16 standardized compute trays, each housing 4 TPU chips, and utilizes a 100% liquid cooling system, supporting up to 9216 TPU chips in a single cluster [3][12]. Group 2: Market Size Projections - The AI PCB market is projected to reach $21.65 billion by 2027, driven by increased demand from Google’s TPU shipments and product iterations [4][38]. - The AI liquid cooling market is expected to grow to $20.18 billion by 2027, as the TDP of chips increases, necessitating more efficient cooling solutions [4][39]. - The AI power supply chip market is forecasted to reach $18.39 billion by 2027, influenced by the power architecture changes introduced with TPUv7 [4][41]. Group 3: Component Value Breakdown - The value breakdown for TPU v7 components includes approximately $54,400 for TPU, $4,000 for PCB, $7,000 for liquid cooling, and $7,100 for power supply, totaling around $73,000 per TPU unit [4][12]. - The projected market sizes for AI PCB, liquid cooling, and power supply chips by 2027 are $36.9 billion, $60.6 billion, and $31 billion respectively, based on Google’s procurement estimates [4][41]. Group 4: Technological Innovations - TPU v7 introduces a dual-chiplet design that integrates two logic cores with eight HBM3e memory stacks, achieving a memory bandwidth of 7.4 TB/s and a peak performance of 4614 TFLOPS [10][33]. - The TPU v7 architecture employs a high-voltage direct current (HVDC) power supply system, which significantly reduces transmission losses and enhances efficiency [12][30]. Group 5: Cooling and Power Supply Innovations - The TPU v7 utilizes a fully liquid-cooled architecture with advanced flow control systems to manage heat dissipation effectively, ensuring stable operating temperatures [25][26]. - The power supply architecture for TPU v7 is designed to handle over 100 kW per cabinet, utilizing a distributed architecture that minimizes current transmission losses [30][32].