CarbonClarity: Understanding and Addressing Uncertainty in Embodied Carbon for Sustainable Computing

TL;DR

CarbonClarity addresses the challenge of uncertain embodied carbon in ICT hardware by introducing a probabilistic framework that models EPA, GPA, yield, and CI_fab as distributions using kernel density estimation, enabling end-to-end uncertainty analysis and safer, more informed design choices. By integrating with ACT and demonstrating across technology nodes, the approach reveals up to a $1.6\times$ gap between mean and a $95^{th}$ percentile carbon estimate for the 7nm node, and shows chiplet architectures can yield substantial reductions in both mean carbon and its uncertainty. The framework supports uncertainty-aware provisioning (e.g., TPU scaling) and exposes key uncertainty drivers, notably EPA and yield, while suggesting renewables to mitigate EPA variability. Collectively, these results enable robust, carbon-aware hardware design and procurement decisions that balance performance and sustainability under real-world variability.

Abstract

Embodied carbon footprint modeling has become an area of growing interest due to its significant contribution to carbon emissions in computing. However, the deterministic nature of the existing models fail to account for the spatial and temporal variability in the semiconductor supply chain. The absence of uncertainty modeling limits system designers' ability to make informed, carbon-aware decisions. We introduce CarbonClarity, a probabilistic framework designed to model embodied carbon footprints through distributions that reflect uncertainties in energy-per-area, gas-per-area, yield, and carbon intensity across different technology nodes. Our framework enables a deeper understanding of how design choices, such as chiplet architectures and new vs. old technology node selection, impact emissions and their associated uncertainties. For example, we show that the gap between the mean and 95th percentile of embodied carbon per cm$^2$ can reach up to 1.6X for the 7nm technology node. Additionally, we demonstrate through case studies that: (i) CarbonClarity is a valuable resource for device provisioning, help maintaining performance under a tight carbon budget; and (ii) chiplet technology and mature nodes not only reduce embodied carbon but also significantly lower its associated uncertainty, achieving an 18% reduction in the 95th percentile compared to monolithic designs for the mobile application.

Figures (13)

• Figure 1: CarbonClarity incorporates sources of uncertainty into embodied carbon modeling to generate probabilistic distributions of quantitative embodied carbon footprints. In contrast, prior tools such as ACTact and GreenChipgreenchip provide deterministic point estimates without offering quantitative bounds on uncertainty. FOCAL focal, on the other hand, considers uncertainty through a comparative analytical model that constructs a uniform uncertainty range but relies on designers having prior knowledge of whether the design is operational or embodied carbon dominated. By modeling uncertainty explicitly, CarbonClarity provides quantitative statistical insights that can more accurately guide computing system and hardware design compared to prior works actgreenchipfocal.
• Figure 2: An overview of the probabilistic modeling framework. The embodied carbon of the SoC is influenced by uncertainties in yield, CI$_{fab}$, EPA, and GPA, captured in various spatio-temporal formats. These uncertainty data are extracted from publicly available reports to construct the probabilistic modeling framework.
• Figure 3: Various greenhouse gases distribution for 14nm technology node
• Figure 4: The process of constructing a probability density using KDE for 16nm technology node based on its EPA value from 2015 to 2019. The left plot shows individual Gaussian kernels generated from each temporal EPA data point. The right plot displays the resulting probability density curve, obtained by averaging the sum of all kernels.
• Figure 5: The left figure shows the decrease of defect density per cm$^2$ for 5nm, 7nm and 10nm nodes since mass production defectDensity. The right figure shows two 10nm chips with different areas can exhibit varying yields over time.