Source-linked AI summary
Orbital AI Computing: Carbon Tradeoffs Across Satellite Scale
Nisha Sarwar, Lei Jiang, Fan Chen
TL;DR
Orbital AI computing lacks accelerator-aware sustainability estimates, despite wide variation in hardware power, mass, and compute characteristics. This paper extends ESpaS with accelerator-aware modeling and finds that hardware choice materially affects the space–ground carbon tradeoff, with larger systems achieving lower energy-normalized carbon intensity through launch-cost amortization.
Problem
Existing ESpaS estimates use generic datacenter hardware, limiting lifecycle carbon comparisons for diverse modern AI accelerators.
Method
The study extends and reimplements ESpaS with accelerator profiles parameterized by power, compute throughput, and system mass, evaluating Jetson AGX Orin and DGX H100 systems.
Results
197.77 vs. 266.45 gCO2e/kWh: DGX H100 achieves lower orbital energy-normalized carbon intensity than Jetson AGX Orin despite higher absolute launch emissions.
Takeaways & Limitations
The space–ground carbon tradeoff is highly sensitive to hardware characteristics, supporting accelerator-aware baselines for orbital AI computing.
Takeaways & Limitations
The study evaluates only two accelerator profiles and partially models embodied carbon, while assuming full hardware utilization and approximating Orin die area.
Abstract
from arXiv · showhide
Low Earth Orbit (LEO) computing is emerging for low-latency, globally distributed AI services, enabled by advances in satellite constellations and reusable launch systems. However, its sustainability remains unclear. Prior work introduces ESpaS, a framework for estimating lifecycle carbon intensity, but models systems using generic datacenter configurations and does not capture modern AI hardware, where power, mass, and compute characteristics vary widely and launch emissions scale with system mass. In this work, we extend ESpaS with accelerator-aware modeling and evaluate two representative systems: a lightweight Jetson AGX Orin for small satellites and a high-performance DGX H100 enabled by large-payload launch platforms. We show that launch emissions act as a fixed carbon overhead: low-mass systems minimize absolute emissions, while high-performance systems amortize this cost more effectively, reducing carbon intensity. Consequently, the space-ground tradeoff is highly sensitive to hardware choice, highlighting the need for accelerator-aware baselines in orbital AI computing.
1 Introduction
LEO computing is emerging for low-latency, globally distributed services, but existing ESpaS modeling does not capture modern AI accelerators. This work extends ESpaS with accelerator-aware profiles and examines how hardware choice affects the space–ground carbon tradeoff across two representative systems.
- 1 Introduction: LEO computing is emerging as a paradigm for low-latency, globally distributed services enabled by satellite constellations and reusable launch systems.Satellites can serve as distributed compute nodes for applications ranging from data preprocessing to AI inference.
- 1 Introduction: ESpaS uses generic datacenter configurations, limiting its representation of specialized AI accelerators with diverse power, mass, and compute characteristics.The limitation matters because many emerging orbital workloads rely on specialized accelerators.
- 1 Introduction: The study extends ESpaS with accelerator-aware hardware profiles reimplemented in Python and evaluates Jetson AGX Orin and DGX H100 systems across scales.The designs represent a lightweight Jetson AGX Orin system for small satellites and a high-performance DGX H100 system.
- 1 Introduction: Hardware choice materially affects the space–ground carbon gap across DGX H100 and Jetson AGX Orin.The cross-scale case study focuses on how hardware choice influences the space–ground carbon tradeoff.
2 Background and Related Work
LEO computing spans constrained CubeSats using lightweight edge accelerators to heavier platforms supporting high-performance accelerators. ESpaS provides lifecycle carbon modeling but lacks grid-powered terrestrial comparison and accelerator-aware hardware modeling, despite launch and re-entry emissions dominating its footprint.
- LEO Computing Systems: CubeSats are typically a few kilograms and constrained in power, thermal capacity, and volume, making them suitable for Jetson-class edge accelerators and on-device inference.Higher-end launch platforms enable heavier payloads, including large satellites or hosted platforms supporting high-performance accelerators.
- Related Work: ESpaS estimates orbital-computing carbon intensity by combining launch and embodied, operational, and re-entry emissions.For terrestrial systems, ESpaS sets launch to None while retaining a solar- and battery-based configuration.
- Limitations: ESpaS’s terrestrial baseline lacks a grid-powered pathway, limiting comparability with modern datacenters using grid electricity and optimized infrastructure.
- Limitations: ESpaS assumes generic datacenter hardware and does not capture modern AI accelerators, whose power, mass, and compute density vary widely.A DGX H100 node draws ∼10.2 kW at ∼130 kg.
- Limitations: Launch and re-entry emissions dominate ESpaS’s total footprint, so embodied emissions scale with system mass and act as a fixed cost.
3 Methodology
The methodology reimplements ESpaS and extends it with accelerator-aware system modeling, using power, compute throughput, and mass to compare lifecycle carbon impacts across scales. Carbon intensity is evaluated through absolute, energy-normalized, and compute-normalized metrics.
- ESpaS reimplementation: ESpaS is reimplemented in Python with its original SystemConfig interface, reproducing reported results within numerical precision.This validates the implementation as a foundation for subsequent extensions.
- Accelerator-aware modeling: The accelerator-aware abstraction parameterizes each system by peak power draw, peak compute throughput, and system mass.Power determines operational emissions, throughput supports normalization, and mass influences embodied emissions from launch and manufacturing.
- Carbon-intensity metrics: The framework compares systems using absolute lifecycle emissions, energy-normalized intensity (gCO2e/kWh), and compute-normalized intensity (per TOPS).DGX throughput is converted from PFLOPS to TOPS using 1 PFLOP = 1000 TOPS.
4 Case Study
The case study shows that hardware choice strongly shapes orbital AI’s carbon tradeoff: launch emissions scale with mass, while compute density and energy efficiency affect how effectively this fixed overhead is amortized. Although DGX H100 has higher absolute emissions, Jetson AGX Orin has slightly lower compute-normalized emissions, and both orbital systems remain more carbon-intensive than terrestrial baselines.
- Impact of hardware choice: 22,554 kgCO2e for DGX H100 versus 150.9 kgCO2e for Jetson AGX Orin shows launch and re-entry emissions closely reflect system mass.The corresponding system masses are 130.45 kg and 0.87 kg, respectively.
- Compute- and energy-normalized metrics: Jetson AGX Orin has slightly lower emissions per unit of compute, whereas DGX H100 has lower carbon intensity per unit energy through improved amortization of fixed launch costs.Launch emissions therefore create a fixed carbon overhead whose effect differs across compute and energy-normalized metrics.
- Compute- and energy-normalized metrics: 34.04 to 161–266 gCO2e/kWh marks the increase in energy intensity from terrestrial baselines to orbital systems.The passage states that orbital systems remain substantially more carbon-intensive than terrestrial baselines.
- Key insight: System mass, power consumption, and compute density jointly determine how effectively launch emissions are amortized.These results motivate accelerator-aware modeling for sustainable AI in LEO.
5 Limitations and Future Work
The study’s conclusions are limited by its narrow accelerator coverage and partial embodied-carbon accounting. It evaluates only DGX H100 and Jetson AGX Orin, while omitting TPU-class systems and several system-level manufacturing components.
- Limitations: The evaluation covers only two accelerator profiles, DGX H100 and Jetson AGX Orin, rather than the full spectrum of AI hardware.The passage identifies these as representative of edge and high-performance systems.
- Limitations: TPU-class systems are not included in the accelerator evaluation.This omission is identified as a specific gap in hardware coverage.
- Limitations: Embodied-carbon modeling includes die-level manufacturing and launch/re-entry emissions but omits packaging, chassis, power delivery, and assembly.Die-level manufacturing is modeled via ESpaS CPA, whereas several system-level components are excluded.