GIF
animated eco-friendly FPGA
deciding on workflow
fpga and diligent analog display
different model accuracies and inference time
FPGA synthesized block diagram

EcoLogic 🌍

Inspiration ⚡️

With data centers consuming massive amounts of power and resources, projections indicate that, without significant efficiency improvements, we risk depleting global resources by 2040. Data centers now contribute around 3% of global energy consumption [1], with that percentage rising alongside AI-driven demand, creating a “carbonivore” effect [2] through increased water and energy use. In addition, AI’s rapid expansion in data centers amplifies the strain on local resources and further stresses ecosystems already impacted by climate change.

In this critical context, EcoLogic presents a new approach: a highly efficient AI solution to drastically reduce data center energy consumption while accelerating inference speed for real-time disaster response. [3]

What EcoLogic Does 🌐

EcoLogic is a breakthrough in energy-efficient AI processing. Using the Differentiable Logic Networks [4], it achieves rapid, real-time inference speeds with minimal power consumption. With our deepest model, our implementation on the DE10-Lite FPGA demonstrates that EcoLogic performs inference in an average of 239 microseconds on the same model which an NVIDIA A100 (a GPU commonly used in large data centers) averages 1087 microseconds. This shows that our approach achieves an increased performance of 78% in comparison to current approaches. This increase in performance continues to increase with larger, deeper models, modeling the scalability of Differentiable Logic Networks.

How We Built It 🛠

Our team brought together students majoring in data science, electrical engineering, computer engineering. Matheus coded and tested 25 different DiffLogic models of varying sizes. Landon handled the hardware by converting the learned logic gates into FPGA-compatible Verilog and VHDL code and conducted timing analysis for FPGA metrics. Stephen, conducted a preliminary literature review to assess the project feasibility, assisting with design, workflow, and research. Raul implemented the Watson RAG model generator, as well as the Watson-based Hurricane information system.

EcoLogic’s DiffLogic architecture employs innovative, differentiable logic gates that support backpropagation across 16 gate types, such as AND, OR, and NOT, without relying on traditional weight-based connections [make this sound better, mention real-valued logic and how reg logic gate networks are not differentiable]. During inference, only the gate with the highest likelihood for each node is activated, resulting in a sparsely connected, highly efficient model optimized for real-time applications. DiffLogic already achieves up to 4x faster inference than standard neural networks. However, by transferring this learned logic gate architecture onto an FPGA, a first in the field to the authors' knowledge, EcoLogic accelerates inference even further, achieving unprecedented speed and efficiency ideal for immediate decision-making in natural disaster scenarios.

Why IBM Should Take Notice 💼

IBM’s leadership in neurosymbolic AI aligns closely with EcoLogic’s design philosophy. By combining symbolic reasoning and neural networks, IBM is already positioned as a pioneer in efficient AI. EcoLogic enhances this mission, increasing the inference speed for AI applications while addressing climate-related challenges by cutting down on the demand for large, power-hungry data centers. This is critical as studies project a rise in both data center energy costs and their carbon impact over the next decade due to increasing demand from AI workloads [5][6].

As a potential IBM integration, EcoLogic could directly support IBM’s AI solutions, enhancing IBM’s offerings to clients needing faster, more sustainable AI. With its ability to perform high-speed, energy-efficient processing, EcoLogic offers IBM a way to lower operational costs and carbon emissions in AI-powered data centers.

Challenges We Overcame 🚧

Developing an FPGA-optimized AI model required addressing technical challenges like hardware mapping and power consumption. Traditional hardware, such as CPUs and GPUs, struggles to match the power efficiency of FPGAs. Adapting DiffLogic for an FPGA without compromising accuracy took extensive experimentation, especially in balancing inference speed and memory constraints. Implementing algorithms such as argmax at an optimal hardware level presented an array of challenges to be conquered. These efforts allowed us to pioneer a model for efficient, AI-based disaster response on hardware.

Accomplishments We’re Proud Of 🎉

Our FPGA implementation is implemented completely by using combinational logic. Without the presence of any registered values, there are no clocked gates to slow down the inference speed, meaning the only limitation for the performance of the FPGA is the propagation delay of the synthesized gates. The scaling of this propagation delay is much lower than a traditional GPU implementation, which significantly speeds up the inference time. Even with the largest model we tested, only 17% of the FPGA's resources were used, meaning that there is still space for larger, more accurate models to be implemented if desired.

What We Learned 📘

EcoLogic underscored the importance of interdisciplinary collaboration, where insights from AI, hardware design, and sustainability perspectives helped refine and achieve our goals. Additionally, we gained hands-on experience in neurosymbolic AI, hardware acceleration, and sustainable AI applications.

What’s Next for EcoLogic 🌱

EcoLogic aims to extend beyond natural disaster preparedness into sectors where real-time, low-power AI is vital, such as environmental monitoring and smart grid management. Our goal is to continue testing on larger datasets and investigate additional FPGA models for broader scalability.

References 📚

IDC Blog. "Data Centers and Our Climate." IDC, 2023. Link
MIT Press Reader. "The Staggering Ecological Impacts of Computation and the Cloud." MIT Press, 2023. Link
McKinsey. "Data centers and AI: How the energy sector can meet power demand." McKinsey, 2023. Link
Petersen, F., Borgelt, C., Kuehne, H., & Deussen, O. "Deep Differentiable Logic Gate Networks." NeurIPS 2022. arXiv preprint. Link
IBM Research. "FPGA-Based Near-Memory Acceleration of Modern Data-Intensive Applications." IEEE Micro, 2021. Link
IBM Research. "Measuring and Modeling the Power Consumption of Energy-Efficient FPGA Coprocessors for GEMM and FFT." Journal of Signal Processing Systems, 2015. Link

Built With

de10-lite
difflogic
ibm-watson
python
quartus
verilog
vhdl
waveforms

Submitted to

UF/IBM AI Days Hackathon Supported by Swamphacks! https://discord.gg/4EJhabM79p
- Winner Most Popular

Created by

I worked on the presentation and the video for submission, prepared documentation, work on data preprocessing, and helped train the models and analyze results

swormald Wormald
I integrated Watson into the system to warn users based on the real-time system outputs of the FPGA. The prompt.py script supports automatically flashing quartus files onto the DE-10 Lite and implementing new RAG models for the user's explainable-logic needs. The dad_analysis.py script is our hurricane sampling solution which notifies users about the Hurricane threat by interfacing with the DAD3 Logic Analyzer.

Raul Valle
I trained the differentiable logic gate models and converted the learned logic gates to verilog and vhdl

Matheus Kunzler Maldaner
I like Neurosymbolic AI
I implemented the generated verilog models and configured the FPGA to synthesize the digital design. I verified all HDL implementations with testbenches. I also ran timing analysis on the FPGA to determine the efficiency compared to traditional GPU implementations.

Landon Nayab