Skip to main content

Edge AI Hardware Optimization

A reference pipeline for evaluating compact CNN deployments under edge-device constraints. Optimize models through pruning, quantization, and hardware-aware analysis.

Get StartedAPI Reference

Quick Start

Get up and running with the optimization pipeline in minutes

1

Install dependencies

Create a virtual environment and install the required packages:
python -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt
The pipeline requires PyTorch, torchvision, matplotlib, pandas, PyYAML, ONNX, and ONNXRuntime.
2

Configure your experiment

Set the Python path and review the default configuration:
export PYTHONPATH=src
The configs/default.yaml file contains deterministic baseline settings including pruning levels, precision modes, and memory budgets.
seed: 7
pruning_levels: [0.0, 0.25, 0.5, 0.7]
precisions: [fp32, fp16, int8]
memory_budgets_mb: [1.0, 2.0, 4.0]
active_memory_budget_mb: 2.0
3

Run the pipeline

Execute the complete optimization pipeline:
python scripts/run_pipeline.py --config configs/default.yaml
This will train a baseline CNN, sweep through pruning and precision variants, and generate Pareto frontiers.
4

Analyze results

The pipeline generates comprehensive outputs in the outputs/ directory:
  • sweep_results.csv — All model variants with metrics
  • pareto_frontier_latency.csv — Optimal latency-accuracy tradeoffs
  • pareto_frontier_energy.csv — Optimal energy-accuracy tradeoffs
  • hardware_summary.csv — Bandwidth utilization and compute estimates
  • Visualization plots for accuracy vs latency, energy, and memory
For production-grade claims, run multiple seeds and aggregate results externally for statistical confidence.

Key Features

Hardware-aware optimization tools for edge AI deployment

Structured Pruning

Remove whole channels from convolutional layers to reduce model size while preserving dense kernel compatibility.

Multi-Precision Support

Evaluate FP32, FP16, and INT8 variants with calibration-based quantization for optimal performance.

Memory Budget Constraints

Enforce SRAM-style memory limits and filter infeasible candidates before Pareto analysis.

Pareto Frontier Analysis

Generate optimal tradeoff curves for latency-accuracy and energy-accuracy to guide deployment decisions.

Layer-wise Profiling

Analyze activation memory, parameter footprints, and MAC operations per layer to identify bottlenecks.

Deterministic Benchmarking

Reproducible latency measurements with configurable benchmark windows and statistical reporting.

Explore by Topic

Deep dive into optimization techniques and hardware analysis

Architecture

Understand the pipeline stages from configuration to Pareto frontier generation.

Model Optimization

Learn how pruning and quantization affect model accuracy and resource usage.

Hardware Constraints

Explore memory budgets, bandwidth utilization, and CPU frequency scaling.

Configuration Guide

Customize experiment parameters including datasets, batch sizes, and benchmarking settings.

Bandwidth Utilization

Estimate achieved bandwidth and identify compute vs transfer bottlenecks.

Precision Tradeoffs

Compare mean accuracy, latency, and memory across FP32, FP16, and INT8 modes.

Ready to optimize your models?

Start with the quickstart guide to run your first optimization sweep, or explore the API reference to integrate the pipeline into your workflow.

View QuickstartExplore API