NEXUS-ML

NEXUS: A Novel Energy eXamination Using Switching-activity-based Tokens

A physics-grounded energy efficiency framework for machine learning systems.

⚠️ Pre-alpha: This package is under active development as part of PhD research.

Overview

NEXUS moves beyond traditional FLOP-based energy proxies to transistor-level energy modeling derived from semiconductor circuit physics. It provides a complete pipeline for understanding, measuring, and optimizing the energy consumption of ML models.

The Problem

FLOPs are a poor proxy for energy consumption. A single memory access costs 100× more energy than a compute operation, yet FLOP counts treat all operations equally. This leads to misleading efficiency comparisons and misguided optimization efforts.

The NEXUS Approach

NEXUS uses β-coefficients derived from Horowitz (ISSCC 2014) semiconductor data to decompose model energy into transistor operations, then computes six novel metrics that reveal where energy actually goes:

Pipeline

Analyze → Measure → Diagnose → Recommend → Optimize

1. Analyze: Profile a model — compute FLOPs (baseline) and transistor operations (NEXUS)
2. Measure: Collect actual hardware energy measurements (GPU via NCU, CPU power monitors)
3. Diagnose: Identify energy bottlenecks at the layer/operation level
4. Recommend: Propose targeted optimizations ranked by expected energy impact
5. Optimize: Apply optimizations (quantization, pruning, activation swaps) guided by NEXUS metrics

Installation

# Core (metrics computation only)
pip install nexus-ml

# With PyTorch support
pip install nexus-ml[pytorch]

# With hardware measurement support
pip install nexus-ml[measurement]

# Everything
pip install nexus-ml[all]

Quick Start

import nexus_ml

# Coming soon — API is under active development

Supported Backends

ML Frameworks

• PyTorch — Full support (CNNs, RNNs, Transformers)
• XGBoost — Gradient boosted decision trees
• scikit-learn — Traditional ML models
• Extensible — Implement ModelBackend for custom frameworks

Hardware

• NVIDIA GPU — Via Nsight Compute (NCU) and pynvml
• CPU — Via system power monitoring
• Extensible — Implement HardwareBackend for custom hardware

Development

# Clone the repository
git clone https://github.com/jemsbhai/nexus-ml.git
cd nexus-ml

# Create virtual environment
python -m venv .venv
.\.venv\Scripts\Activate.ps1

# Install in development mode with all dependencies
pip install -e ".[all]"

# Run tests
pytest tests/ -v

# Run tests with coverage
pytest tests/ -v --cov=nexus_ml --cov-report=term

Project Structure

nexus-ml/
├── src/nexus_ml/          # Main package
│   ├── core/              # β-coefficients, TO counting, energy model
│   ├── metrics/           # NEXUS metrics + baseline (FLOPs) metrics
│   ├── analyze/           # Model profiling & decomposition
│   ├── measure/           # Hardware energy measurement
│   ├── diagnose/          # Bottleneck identification
│   ├── recommend/         # Optimization suggestions
│   ├── optimize/          # Apply optimizations
│   ├── backends/          # ML framework adapters (pluggable)
│   ├── hardware/          # Hardware adapters (pluggable)
│   └── report/            # Visualization & export
├── tests/                 # Test suite (TDD)
├── examples/              # Usage examples
├── docs/                  # Documentation
└── packages/              # Future ports (JS/TS)

Citation

If you use NEXUS in your research, please cite:

@inproceedings{syed2026toml,
  title={TOML: Transistor Operations for Machine Learning -- A Physics-Grounded Energy Efficiency Framework},
  author={Syed, Muntaser and Silaghi, Marius},
  booktitle={Proceedings of the Florida Artificial Intelligence Research Society (FLAIRS-39)},
  year={2026}
}

License

MIT License — see LICENSE for details.

Acknowledgments

• β-coefficient data derived from Horowitz, "Computing's Energy Problem (and what we can do about it)", ISSCC 2014
• PhD research conducted at Florida Institute of Technology under Dr. Marius Silaghi