Latest News

SustainML has successfully integrated FPGA-based hardware estimation into its AI sustainability evaluation pipeline, enabling more accurate energy and carbon footprint analysis for CNN models deployed on FPGA platforms. By combining hardware-level latency and power prediction with automated carbon intensity calculation, the framework now supports end-to-end sustainability assessment for both Hugging Face transformer models and ONNX-based CNN models targeting FPGA devices. This integration strengthens SustainML’s ability to evaluate AI models across heterogeneous hardware environments and brings hardware-aware carbon estimation closer to real-world deployment scenarios.

 SustainML continues to advance its mission of enabling sustainable and hardware-aware artificial intelligence by successfully integrating FPGA-based energy and carbon estimation into its evaluation pipeline. This milestone marks an important step toward providing accurate, end-to-end sustainability assessment for AI models across heterogeneous computing platforms.

As AI workloads increasingly move beyond traditional CPU and GPU environments into specialized accelerators such as FPGAs and processing-in-memory architectures, it becomes essential to evaluate not only model accuracy and performance, but also their environmental impact. SustainML addresses this need by combining model profiling, hardware-specific power estimation, and carbon footprint computation within a unified framework.

With this latest development, SustainML now supports CNN models deployed in ONNX format on FPGA platforms, alongside its existing support for Hugging Face transformer models evaluated through simulation-based hardware profiling.

Hardware-Aware Energy Estimation for FPGA Platforms

The integration introduces FPGA-based latency and power estimation directly into the SustainML workflow. When a CNN model targeting FPGA hardware is selected, the framework leverages a dedicated FPGA predictor capable of estimating:

• Execution latency
• Power consumption
• Hardware-specific performance characteristics

These values are then translated into energy consumption metrics using standard physical relationships between power, time, and energy. From there, carbon emissions are computed using an appropriate grid carbon intensity factor.

This approach ensures that the carbon footprint estimation is grounded in realistic hardware behavior rather than purely theoretical assumptions. By using hardware-aware predictors, SustainML can now provide meaningful sustainability insights for FPGA-targeted AI workloads.

Unified Carbon Estimation Across Hardware Types

A key achievement of this integration is the harmonization of carbon footprint calculation across different hardware backends. SustainML now supports:

• Transformer-based models evaluated via profiling and runtime energy tracking
  ONNX-based CNN models targeting FPGA accelerators
• Simulation-based architectures such as processing-in-memory systems

Despite the diversity of hardware paths, the framework applies a consistent methodology:

1. Measure or estimate latency and power consumption
2. Convert these into energy consumption (kWh)
3. Apply grid carbon intensity to compute CO₂ emissions
4. Report per-inference energy and carbon metrics

This unified approach ensures that results are comparable across different hardware platforms and model families, enabling informed decisions about sustainability trade-offs.

Flexible Architecture and Robust Design

The FPGA integration was designed to fit seamlessly into SustainML’s modular node-based architecture. The Hardware Resources node selects the appropriate backend (FPGA predictor or simulator), while the Carbon Footprint node computes final sustainability metrics.

This separation of concerns ensures:

• Maintainability
• Extensibility to additional hardware platforms
• Clear data flow between hardware estimation and carbon computation

The framework remains adaptable for future integrations, including additional accelerators, improved carbon intensity data sources, or region-specific sustainability parameters.

Strengthening Sustainable AI Engineering

This development reinforces SustainML’s role as a practical tool for sustainable AI engineering. By incorporating FPGA-based carbon estimation, the platform moves closer to real deployment environments where AI models increasingly rely on specialized accelerators for performance and efficiency.

As sustainability becomes a core requirement in AI system design, tools like SustainML play a critical role in enabling measurable, transparent, and reproducible carbon assessments.

Looking Ahead

The FPGA integration lays the foundation for further advancements in hardware-aware carbon modeling. By continuously extending its hardware coverage and refining its carbon estimation methodology, SustainML remains committed to advancing sustainable AI practices across the full stack—from model design to hardware deployment.

This milestone represents another important step toward making environmental impact a first-class metric in AI development. Through FPGA integration and unified carbon accounting, SustainML strengthens its capability to support responsible, energy-efficient artificial intelligence across diverse hardware ecosystems.