About the work
This registered work comprises the source code and hardware–software implementation of the GMVigIA On-Board Artificial Intelligence convolutional neural network, specifically designed for on-board Earth Observation (EO) data processing on AMD Xilinx Versal ACAP platforms.
The implementation is based on an optimized and fully quantized INT8 convolutional encoder–decoder neural network, tailored to meet the stringent performance, power, and resource constraints of spaceborne EO missions. The network architecture enables efficient feature extraction, compression, and reconstruction for advanced EO applications such as onboard image enhancement, semantic analysis, and data reduction.
The registered material includes:
An INT8-quantized convolutional encoder–decoder network, optimized for inference efficiency while preserving EO data fidelity.
AI Engine-ML (AIE-ML) kernels implementing convolutional layers and activation functions, designed for parallel, vectorized execution across multiple AIE tiles to maximize throughput and energy efficiency.
RTL kernels in VHDL for supporting neural network operations, including data reorganization, MaxPooling, upsampling, and feature map concatenation, offloading non-convolutional processing from the AIEs and enabling efficient pipelined execution.
A data-driven control and buffering infrastructure for Versal ACAP, implementing flow control, synchronization, and high-throughput data movement between Programmable Logic (PL), AI Engines, and memory interfaces.
The overall architecture leverages the heterogeneous computing capabilities of Versal ACAP, combining AI Engines, programmable logic, and adaptive dataflow to deliver a scalable, modular, and space-ready on-board AI processing solution. The source code reflects design choices aimed at radiation-tolerant adaptation, deterministic execution, and future extensibility, supporting European technology independence in advanced EO missions.
AI Availability Declaration
This work cannot be made available to AI systems.
Creativity declaration
No AI has been used in the creative process of this work
Print work information
Work information
Title GMVigIA On-Board AI Convolutional Network for Earth Observation on Versal ACAP
This registered work comprises the source code and hardware–software implementation of the GMVigIA On-Board Artificial Intelligence convolutional neural network, specifically designed for on-board Earth Observation (EO) data processing on AMD Xilinx Versal ACAP platforms.
The implementation is based on an optimized and fully quantized INT8 convolutional encoder–decoder neural network, tailored to meet the stringent performance, power, and resource constraints of spaceborne EO missions. The network architecture enables efficient feature extraction, compression, and reconstruction for advanced EO applications such as onboard image enhancement, semantic analysis, and data reduction.
The registered material includes:
An INT8-quantized convolutional encoder–decoder network, optimized for inference efficiency while preserving EO data fidelity.
AI Engine-ML (AIE-ML) kernels implementing convolutional layers and activation functions, designed for parallel, vectorized execution across multiple AIE tiles to maximize throughput and energy efficiency.
RTL kernels in VHDL for supporting neural network operations, including data reorganization, MaxPooling, upsampling, and feature map concatenation, offloading non-convolutional processing from the AIEs and enabling efficient pipelined execution.
A data-driven control and buffering infrastructure for Versal ACAP, implementing flow control, synchronization, and high-throughput data movement between Programmable Logic (PL), AI Engines, and memory interfaces.
The overall architecture leverages the heterogeneous computing capabilities of Versal ACAP, combining AI Engines, programmable logic, and adaptive dataflow to deliver a scalable, modular, and space-ready on-board AI processing solution. The source code reflects design choices aimed at radiation-tolerant adaptation, deterministic execution, and future extensibility, supporting European technology independence in advanced EO missions.
Work type Source Code
Tags convolutional-neural-network, edge-ai, versal-acap, vhdl, int8-quantization, encoder-decoder, earth-observation, rtl-kernels, space-applications, aie-ml, onboard-ai, gmv
-------------------------
Registry info in Safe Creative
Identifier 2512234099654
Entry date Dec 23, 2025, 11:28 AM UTC
License All rights reserved
-------------------------
Copyright registered declarations
Author. Holder GMV Aerospace and Defence SAU. Date Dec 23, 2025.
Information available at https://www.safecreative.org/work/2512234099654-gmvigia-on-board-ai-convolutional-network-for-earth-observation-on-versal-acap
/
Registration: 2512234099654
