Life-Cycle Emissions of AI Hardware: A Cradle-To-Grave Approach and Generational Trends

TL;DR

This paper tackles the problem of quantifying cradle-to-grave greenhouse gas emissions for AI hardware by performing a comprehensive life-cycle assessment (LCA) of Google's TPU accelerators, including the first manufacturing-emissions estimates for AI accelerators. It introduces Compute Carbon Intensity (CCI), a normalized metric defined as CO2e per FLOP, and decomposes it into embodied and operational components measured with real fleet data and a six-year functional unit. The study finds that operational emissions dominate lifecycle emissions (roughly 70–90%), while manufacturing emissions rise with newer generations but overall CCI improves threefold from TPU v4i to v6e due to throughput gains and efficiency gains, with potential further reductions under 24/7 carbon-free energy scenarios. The work provides a detailed methodology, comparisons with corporate inventories, and workload-level analyses, offering actionable guidance for hardware-software optimization and industry-standard LCAs that can inform design, procurement, and policy.

Abstract

Specialized hardware accelerators aid the rapid advancement of artificial intelligence (AI), and their efficiency impacts AI's environmental sustainability. This study presents the first publication of a comprehensive AI accelerator life-cycle assessment (LCA) of greenhouse gas emissions, including the first publication of manufacturing emissions of an AI accelerator. Our analysis of five Tensor Processing Units (TPUs) encompasses all stages of the hardware lifespan - from raw material extraction, manufacturing, and disposal, to energy consumption during development, deployment, and serving of AI models. Using first-party data, it offers the most comprehensive evaluation to date of AI hardware's environmental impact. We include detailed descriptions of our LCA to act as a tutorial, road map, and inspiration for other computer engineers to perform similar LCAs to help us all understand the environmental impacts of our chips and of AI. A byproduct of this study is the new metric compute carbon intensity (CCI) that is helpful in evaluating AI hardware sustainability and in estimating the carbon footprint of training and inference. This study shows that CCI improves 3x from TPU v4i to TPU v6e. Moreover, while this paper's focus is on hardware, software advancements leverage and amplify these gains.

Figures (8)

• Figure 1: A comprehensive AI cradle-to-grave emissions inventory boundary. It excludes GHG emissions from edge and end-user consumer devices as they are beyond the AI product developer’s control. It omits auxiliary compute and storage emissions sources given the definition of the accelerator-host hardware functional unit.
• Figure 2: Lifetime AI hardware compute carbon intensity (CCI), with a breakdown by life-cycle stage.
• Figure 3: Accelerator and host emissions by life-cycle stage.
• Figure 4: Absolute manufacturing emissions per chip, by generation and by component type, and average TeraFLOP/second. It shows the total manufacturing emissions per machine, normalized by the number of chips per machine. (Table \ref{['tab:specs']} embodied includes DC construction and transportation emissions.) The memory piece of this CO$_{2}$e (TPU HBM + CPU DRAM) averages $\sim$38%.
• Figure 5: Typical hardware components in a server.