Innervator: Hardware Acceleration for Neural Networks

Pasted below is an overview/abstract, and you will find more information (including a paper, demo video, statistics, slides, and source code) at the following GitHub repository:
https://github.com/Thraetaona/Innervator
------------------------------------------------------------------------
Artificial intelligence ("AI") is deployed in various applications, from noise cancellation to image recognition, but AI-based products often come with high hardware and electricity costs; this makes them inaccessible for consumer devices and small-scale edge electronics. Inspired by biological brains, deep neural networks ("DNNs") are modeled using mathematical formulae, yet general-purpose processors treat otherwise-parallelizable AI algorithms as step-by-step sequential logic.   In contrast, programmable logic devices ("PLDs") can be customized to the specific parameters of a trained DNN, thereby ensuring data-tailored computation and algorithmic parallelism at the register-transfer level. Furthermore, a subgroup of PLDs, field-programmable gate arrays ("FPGAs"), are dynamically reconfigurable.  So, to improve AI runtime performance, I designed and open-sourced my hardware compiler: Innervator.  Written entirely in VHDL-2008, Innervator takes any DNN's metadata and parameters (e.g., number of layers, neurons per layer, and their weights/biases), generating its synthesizable FPGA hardware description with the appropriate pipelining and batch processing. Innervator is entirely portable and vendor-independent.  As a proof of concept, I used Innervator to implement a sample 8x8-pixel handwritten digit-recognizing neural network in a low-cost AMD Xilinx Artix-7(TM) FPGA @ 100 MHz.  With 3 pipeline stages and 2 batches at about 67% LUT utilization, the Network achieved ~7.12 GOP/s, predicting the output in 630 ns and under 0.25 W of power.  In comparison, an Intel(R) Core(TM) i7-12700H CPU @ 4.70 GHz would take 40,000-60,000 ns at 45 to 115 W. Ultimately, Innervator's hardware-accelerated approach bridges the inherent mismatch between current AI algorithms and the general-purpose digital hardware they run on.
------------------------------------------------------------------------