MicroGreen

MicroGreen, a design space exploration framework for sustainable edge devices. MicroGreen integrates embodied-carbon modeling, workload characterization, and environment-dependent operational analysis. It incorporates parametric embodied-carbon models for MCUs, sensors, regulators, storage, batteries, capacitors, and solar panels, leveraging LCA databases. Using these models, MicroGreen takes as input application requirements (e.g., workload, duty cycle, deployment lifetime) and environmental traces (e.g., solar availability) and identifies carbon-optimal configurations. We empirically characterize a repository of MCUs under diverse workloads, integrate peripheral and power-source options, and generate Pareto-optimal designs to illustrate trade-offs between carbon footprint, performance, and cost

---

Setup

1. Clone the repository with all submodules:

git clone --recurse-submodules git@github.com:S4AI-CornellTech/MicroGreen.git
cd MicroGreen

2. Install Python dependencies using the setup script:

bash setup.sh

---

Repository Structure

• framework/: Core MicroGreen framework and Streamlit UI
• profiling/: Hardware profiling code and setup instructions
• database/: Precomputed embodied carbon and profiling results
• EmbodiedCarbonModeling/: Embodied carbon modeling submodule

---

Reproducing Key Result Figures in MicroGren

Figure 3 - Board Carbon Breakdown

python3 scripts/carbon_component_composition_plotter.py

Figure 5 - Per Inference Runtime and Energy Plot

python3 scripts/characterization_fig.py

Figure 6 - Carbon Rank Plot

python3 scripts/overall_eval_carbon.py --lifetime-years 1 --solar-panel-area-cap 611

Figure 7 - Energy Efficiency Rank Plot

python3 scripts/overall_eval_energy.py

Figure 8 - Irradiance Analysis Plot

python3 framework/main.py --workload kws-l --solar-plot

Figure 9 - Battery Analysis Plot

python3 framework/main.py --workload kws-s --battery-plot

Figure 10 - Network Analysis Plot

python3 scripts/network_plot.py

Figure 11 - Hybrid Analysis Plot

python3 framework/main.py --workload kws-l --lifetime-plot

Figure 14 Heterogeneous Deployment Results

python3 framework/heterogeneousDeployment.py
python3 scripts/case_study_plot.py

---

Carbon Modeling Verification

The embodied carbon estimates in database/board_carbon.csv are derived from the EmbodiedCarbonModeling/ submodule. To regenerate the database from the modeling outputs, first follow the submodule's README.md to generate carbon estimates from BoMs, then run the following script to port the results to database/:

python3 scripts/board_carbon_csv_generator.py \
  EmbodiedCarbonModeling/outputs/coralDevMicro_output \
  EmbodiedCarbonModeling/outputs/ESP32_output \
  EmbodiedCarbonModeling/outputs/ESP32-C6_output \
  EmbodiedCarbonModeling/outputs/ESP32-S3_output \
  EmbodiedCarbonModeling/outputs/nRF52840_output \
  EmbodiedCarbonModeling/outputs/RP2040_output \
  EmbodiedCarbonModeling/outputs/RP2350_output \
  EmbodiedCarbonModeling/outputs/STM32F411_output \
  -o database/board_carbon.csv

---

Profiling Verification

The database/ directory contains per-MCU inference latency and power measurements collected using the hardware setup described in the paper. To verify or regenerate these results, refer to the guides below.

Guide	Location	Description
INFERENCE_PROFILING_GUIDE.md	profiling/inference/	Environment setup, compilation, flashing, and latency measurement for all inference workloads running on each MCU
WIRELESS_PROFILING_GUIDE.md	profiling/wireless/	Replicating BLE wireless measurements reported in the paper
POWER_PROFILING_GUIDE.md	profiling/	Hardware, software, and operation details for power profiling

---

License

This project is licensed under the MIT License. See the LICENSE file for full details.