Speaker
Description
The increasing demand for onboard data processing in space
applications has led to the integration of Artificial Intelligence
(AI) and Machine Learning (ML) on Field-Programmable Gate Arrays
(FPGAs). This is particularly relevant for missions requiring
optimized data transmission, such as Earth observation
applications. AI-driven techniques can enhance onboard autonomy by
performing tasks such as event detection, data filtering, and
compression, ultimately reducing downlink bandwidth requirements.
The Edge SpAIce project demonstrates the potential of FPGA-based AI
processing for space applications, focusing on plastic litter
detection in oceans using Deep Neural Networks. Since real-time
inference is not required, our approach prioritizes computational
efficiency, using pixel/second/watt as the primary performance
metric. By balancing latency, throughput, and power consumption, we
optimize FPGA utilization for space-based deployments. Leveraging
open-source tools such as hls4ml and QONNX, we implement drastic
model compression and efficient hardware deployment,
enabling high-performance, low-power computation suitable for
resource-constrained space environments.
Affiliation of author(s)
CERN
| Track | Artificial Intelligence/Machine Learning |
|---|
