The sheer number of artificial intelligence (AI) applications and increasing demand for cloud computing technologies require effective plans for coping with requests related to computing resources. But how should managers of humongous data centers and complex network infrastructures efficiently plan for the future needs of AI applications and cloud computing? In this timely opinion article, Vahdat et al. present the components of a revolutionary, practical, and comprehensive metrics for assessing the environmental impacts and capacity of huge computers in data centers.

The authors compellingly argue that the current performance metrics of computerized systems are inadequate for assessing energy usage and environmental impact. Thus, they present and shed new light on a better understanding of metrics for the comprehensive evaluation of complex computer systems. Indeed, given the rapidly changing technologies, the use of today’s lifetime purchasing costs of system components such as chips can be misleading when assessing total computer system performance.

The article recognizes the limitations and critiques the equation for estimating the total cost of ownership (TCO), defined as the purchasing price with additional operational expenditures of the system’s infrastructure, which depends on the emerging new chips and components. Thus, the authors outline new operational definitions of metrics for estimating the impacts and consequences of power usage and carbon dioxide emissions on the complete system workloads at data centers. The authors succinctly discuss the long-lasting issues of these operational metrics on system workload performance benchmarks.

The article convincingly illustrates, and advocates for, new metrics for evaluating computer systems, with an example that illuminates the effectiveness of two typical tensor processing unit (TPU) AI accelerators. The data analysis from the example clearly calls on data center system administrators to be more cognizant of improved communication speeds and reduced power consumption, while also minimizing the impacts on carbon emission. Consequently, the authors challenge computer system designers to develop new ways to: (1) reduce hardware average power consumption; (2) assess energy usage and reduce and trace carbon emissions; (3) label computer system components with adequate carbon emissions; and (4) minimize carbon emissions associated with computer and semiconductor manufacturing.

Without a doubt, all data center managers should read this practical article and take advantage of the illustrative definitions and equations required to compute practical metrics for effectively sustaining data centers for AI and cloud applications. Computer engineers and researchers will also find the questions raised here mind-boggling.