Artificial intelligence can be remarkably powerful, but putting that intelligence into a small, resource-constrained device is a very different challenge.
For Daniel Gutiérrez, CTO at INTIGIA and winner of the Third Open Call of dAIEDGE, the answer lies in hardware-aware optimization. Through PYNQ-OPT, the team is developing approaches to make complex AI models more efficient for FPGA-based systems, with applications ranging from industrial environments to space.
The project focuses on the fundamental SWaP constraints, Size, Weight and Power, that are particularly important when AI has to operate on embedded hardware. PYNQ-OPT explores how deep learning models can be adapted to fit within these limitations while retaining the accuracy and performance required for real-world applications.
INTIGIA brings extensive experience to this challenge. Over the past years, the team has worked on implementing DSP, computer vision and deep learning algorithms on FPGAs, developing efficient hardware architectures and proprietary methodologies for deploying AI models.
One of the key techniques is quantization: reducing the number of bits used to represent a model while carefully preserving its accuracy. By finding the right balance between model size, computational requirements and performance, the team aims to enable real-time processing on constrained devices.
The project is particularly focused on convolutional neural networks for applications such as real-time vehicle detection, demonstrating how sophisticated computer vision can be brought closer to where data is generated.
Throughout the seven-month programmed, Daniel and the INTIGIA team have worked with German Research Center for Artificial Intelligence, benefiting from its expertise and guidance as hosting institution. We sincerely thank DFKI for supporting the team in turning its FPGA expertise into a new generation of efficient Edge AI solutions.
The dAIEDGE-VLab has also provided an environment to benchmark and refine these approaches, helping the team evaluate how different optimisation strategies perform under real hardware constraints rather than relying solely on theoretical results.
PYNQ-OPT demonstrates a fundamental principle of Edge AI: the best model is not necessarily the biggest or most powerful one, but the one that can deliver the right performance within the hardware constraints of its environment.