中金：谷歌引领ASICs自研加速 异于GPGPU架构的硬件价值再定义

Core Insights - The launch of Google's TPUv7 marks a significant advancement in ASIC clusters, moving away from traditional GPGPU architectures, which is expected to reshape the hardware landscape and accelerate the growth of AI computing hardware markets such as PCB, liquid cooling, and power supply [1] Group 1: Google TPU Evolution - Google has evolved its TPU architecture over the past decade, transitioning from TPU v1 in 2016 to TPUv7, which features nearly 10,000 chip clusters and incorporates OCS optical switching architecture and HBM high-bandwidth memory [2] - The TPUv7 chip has made significant improvements in dual-chip packaging architecture, enhancing linear acceleration in large-scale clusters [2] Group 2: TPUv7 Hardware Changes - The TPUv7 tray architecture includes 16 standardized computing trays, each housing 4 TPU chips; the power architecture utilizes a +/-400V high-voltage direct current (HVDC) solution; and the cooling system employs a 100% liquid cooling architecture with a large cold plate design covering 4 TPU chips and VRM [3] - The maximum cluster size supports interconnection of 144 racks, totaling 9,216 TPU chips [3] Group 3: Market Projections - The breakdown of the TPUv7 cabinet solution indicates the value of TPU, PCB, liquid cooling, power supply, and cables at approximately $54,400, $4,000, $7,000, $7,100, and $400 respectively, totaling around $730,000 [3] - As TPU shipments increase and product structure iterations drive demand, the AI PCB, liquid cooling, and power supply chip markets are projected to grow significantly, with expected market sizes of $36.9 billion, $60.6 billion, and $31 billion respectively by 2027 based on Google's procurement standards; overall GPU and ASICs demand is projected to reach $216.5 billion, $201.8 billion, and $183.9 billion for these markets by 2027 [3]