Aether Engine

A Parameterized CNN Accelerator for FPGAs with Resource-Aware Scaling

Overview

Aether Engine is a versatile CNN accelerator implementation for FPGAs that supports various neural network architectures. The accelerator is designed with flexibility in mind, allowing customization for both resource-constrained and high-performance FPGA targets.

Key Features

• Parameterized design supporting various matrix sizes
• Scalable number of accelerator units
• Configurable for different FPGA resource profiles
• Support for variable input dimensions
• Flexible CNN operation support

Supported Operations

• Convolution layers
• Activation functions
  • Sigmoid
  • ReLU
  • Softmax
• Dense (Fully connected) layers
• Maxpooling

Tested Models

• MNIST (In Progress)
• YOLO (Planned)

Build Instructions

git clone --recurse-submodules https://github.com/fpgabuilds/mnist-fpga.git