ElasticAI Creator: A Neural Network Design, Training, and Code Generation Toolchain Optimized for FPGAs

ElasticAI Creator is an open-source toolchain developed by Germany's University of Duisburg-Essen, designed to support neural network design, training, and compilation into VHDL code for FPGA deployment. It aims to lower the barrier for software developers to deploy AI models on FPGAs by streamlining the end-to-end process from model design to hardware deployment.

As AI applications move to edge devices, running neural networks efficiently on resource-constrained hardware becomes critical. Cloud-based inference has issues like high latency, privacy risks, and network dependency. FPGAs are ideal for edge AI due to customizability, low power, and parallel computing, but deploying models to FPGAs requires hardware expertise and manual optimization, limiting adoption.

ElasticAI Creator is developed by the Intelligent Embedded Systems Lab at the University of Duisburg-Essen, part of the ElasticAI ecosystem. Its core goal is to bridge the gap between neural network design and FPGA deployment, enabling software-background developers to easily deploy AI models on hardware without deep hardware knowledge.

The toolchain follows three key stages:

1. Design & Training: Use elasticai.creator.nn layers optimized for FPGA, including quantization-aware training to preserve model performance after quantization.
2. Intermediate Representation Conversion: Convert trained models to an intermediate representation (IR), then translate to VHDL (with plans to support more hardware description languages).
3. Third-Party Compilation: Compile generated VHDL code using tools like Xilinx Vivado and flash it to target FPGA hardware.

Quantization Support: Deep support for fixed-point, integer-only linear, binary, and limited-precision floating-point quantization; supports layers like Linear (with batch norm), LSTM, Conv1D, and activations (Hard Sigmoid, ReLU, adaptive SiLU, etc.). Supported Architectures: Sequential models (MLP, time-multiplexed networks, GRU) and basic components (counters, sliding windows, grouped filters). VHDL Component Library: Prebuilt components for combinatorial circuits, sequential logic, and neural network-optimized运算 units.

Modular Design:

• creator directory: Core concepts including qtorch (quantized PyTorch layers).
• nn subdirectory: VHDL translation logic for each supported layer type.
• tests directory: Comprehensive unit and integration tests. Elastic Node V5 Integration: Generates VHDL code and C stub code (via elastic-ai.runtime.enV5) to simplify host-FPGA communication.

Limitations:

1. Only supports sequential models (no skip/residual connections).
2. Mainly supports VHDL (other hardware languages under development).
3. No automatic PyPI release (install via Git for latest versions). Future Plans: Optimize fixed-point MAC units, expand convolution layer support, add more network architectures.

Application Scenarios: Industrial IoT (predictive maintenance), smart sensors (privacy-preserving edge AI), academic research (embedded AI experiments), education (software-to-hardware deployment learning). Installation:

• Via pip: python3 -m pip install "elasticai.creator"
• From GitHub: python3 -m pip install git+https://github.com/es-ude/elastic-ai.creator.git@main A minimal example is provided to demonstrate model definition, VHDL translation, and C interface generation.