Data Centers on Wheels: Emissions From Computing Onboard Autonomous Vehicles

Sudhakar, Soumya; Sze, Vivienne; Karaman, Sertaç · 2022 · OpenAlex-citations

DOI: 10.1109/mm.2022.3219803

archive: archived pipeline: cataloged verified

Get this paper ↗ (DOI — opens at the source; we link to it, we don't host it)

Summary

This paper addresses the overlooked environmental impact of computing onboard autonomous vehicles (AVs), characterizing them as "data centers on wheels." While significant attention has been paid to greenhouse gas emissions from stationary data centers, the potential carbon footprint of a global fleet of AVs remains underexplored. The authors aim to quantify these emissions and determine the hardware efficiency improvements required to keep AV computing emissions within acceptable limits relative to current global benchmarks. To achieve this, the authors introduce a probabilistic modeling framework that estimates annual CO2 equivalent emissions based on four key variables: the number of AVs, average computer power per vehicle, average daily driving hours, and the carbon intensity of the electricity used. The model focuses specifically on operational emissions from the autonomy stack’s deep neural network (DNN) workloads, excluding embodied carbon from manufacturing or sensor emissions. The authors parameterize the model using current trends and project future scenarios through 2050. They model AV adoption rates ranging from low to high (up to 95% market share), workload growth (doubling every 3, 5, or 10 years), and hardware energy efficiency improvements based on historical trends of TOPS per Watt. Monte Carlo simulations with one million samples are used to estimate emission distributions under various combinations of these parameters. The results indicate that AV computing emissions have the potential to be comparable to or exceed those of all data centers in 2018. In a high-adoption scenario where 95% of vehicles are autonomous, computer power must remain below 1.2 kW to ensure AV emissions stay below 2018 data center levels in 90% of modeled scenarios. The study finds that maintaining the current rate of hardware efficiency improvement (doubling every 2.8 years) is insufficient to contain emissions if workloads double every three years. To keep 2050 emissions equal to 2018 data center levels under high adoption and business-as-usual decarbonization, hardware efficiency must double every 1.1 years—a rate significantly faster than current trends. Even aggressive decarbonization efforts cannot fully offset emissions if hardware efficiency improvements lag behind workload growth. The significance of this work lies in highlighting the urgent need for specialized hardware and algorithmic efficiency improvements in the AV sector. The authors conclude that general-purpose computing trends are insufficient to manage the carbon footprint of widespread AV adoption. They identify several avenues for future research, including characterizing emissions from sensors, analyzing embodied versus operational carbon over long vehicle lifespans, and exploring hardware specialization for autonomy tasks. The paper calls for industry transparency regarding computer power metrics to enable better assessment and reduction of the carbon footprint associated with autonomous driving.

Provenance

The full processing record for this entry. Every stage of this paper's journey through the pipeline is logged — what ran, with which tool and model, how many attempts it took, and when it last completed.

Summary generated by qwen3.6-27b-prismaquant on 2026-06-26; verification: verified.

Topics

Ranked by relevance to this paper. Hover a topic for its definition.

• driverless ads