Search

Browse

• Communities
• Authors
• Year
• Titles
• Subjects

My Space

• Login

Design and implementation of an FPGA-Based CNN architecture for on-Board satellite processing of data from the Euclid space telescope

Kalomoiris Ioannis

Full record

---

Summary

Convolution Neural Networks (CNNs) have been widely employed for various AI tasks and have demonstrated state-of-the-art performance, especially in complex image recognition problems. The widely used for these tasks GPUs, although having a lot of computational power, come with very high power consumption. This is a deterrent factor for their usage, especially in cases where a small energy footprint is important, like on-board signal processing. In this thesis, we demonstrate an FPGA architecture implemented for the inference stage of a specific CNN, enabling the estimation of the galaxy redshift from spectroscopic observations by dividing the redshift range into 800 Classes. The proposed FPGA architecture achieved an improvement in energy efficiency of up to 11.9x alongside a 2.16x throughput speedup over GPU platforms. The results are from actual executions on FPGAs with space-qualified equivalent parts, enabling performing accurate redshift estimation in space with low energy cost, with no need for raw data transmission to the ground.

Available Files

Services