数据中心建设正如火如荼，算力始终紧缺，存力也存在缺口，而另一个同样重要却未获得足够关注的关键点—— 电 力，也面临着紧张局面。 高盛在一份最新报告中表示，美国在AI领域面临的最大障碍并非芯片、稀土或人才，而是电力。 01 数据中心，需要多少电？ 众所周知，英伟达GPU很耗电。 微软数据中心技术治理和战略部门首席电气工程师曾发布一则数据：以英伟达单个GPU为例，H100 GPU峰值功耗为 700瓦，一小时耗电0.7度，一年按61%的使用时间计算，全年将耗电3740度，相当于一个美国家庭的平均功耗（假设 每个家庭2.51人）。2024年英伟达的H100 GPU销量大约是150万块-200万块，当数百万块 H100部署完毕时，其总功耗 将超过美国亚利桑那州凤凰城所有家庭的用电量。 数据中心的用电量，远远不仅如此。 除了GPU，数据中心中还有大量的设备，比如服务器（还包含CPU等部件）、网络设备、存储设备、冷却系统和照明 等，这些设备无一不需要持续供电。其中数据中心的冷却系统是能耗的主要组成部分之一，总耗电量占到38%以上(有 的甚至高达50%)。传统的空气冷却系统效率较低，而高效的冷却技术，如直接芯片制冷和浸没式液 ...

Core Viewpoint - PowerLattice, a Vancouver-based startup, has developed a micro-solution aimed at addressing the significant energy consumption issues of data centers through a specialized chip technology that enhances power delivery efficiency [2][3]. Group 1: Investment and Leadership - PowerLattice announced it has raised $25 million in new funding for its technology development, with former Intel CEO Pat Gelsinger joining its board [2]. - The investment round also included participation from Playground Global and Celesta Capital, with notable figures from Intel involved [2]. Group 2: Technology and Impact - The company has designed a specialized chiplet that can be installed near processors responsible for AI core computations, allowing for more precise power delivery [2]. - The anticipated outcome is a potential reduction in power consumption by up to 50% while maintaining the same computational capabilities, which could alleviate the energy crisis caused by soaring data center power demands [3][4]. Group 3: Industry Context - The energy crisis is exacerbated by the increasing power demands of AI, leading to significant electricity consumption that can rival that of entire cities [3][4]. - PowerLattice aims to tackle both the energy crisis and the slow development of new computing technologies by providing more efficient computing solutions [4].