Introduction

Klepsydra Technologies set out to enable highly efficient AI inference onboard satellites—for example, enabling real-time image-based geolocation.

Klepsydra collaborated with Enclustra to combine their AI framework with a compact FPGA-based hardware platform to establish a cost-effective and high-performance AI compute platform for satellites. Using the demanding tasks of real-time image-based geolocation as an example, the resulting system demonstrated onboard execution of AI models, with high throughput, minimal CPU load, and stable memory utilization, all within the strict limits of spaceborne systems.

The Challenge

Operating in space means operating without compromise: limited power availability, restricted communication windows, and extreme environmental conditions. Customers need a system that could perform compute-intensive AI inference onboard, deterministically and reliably, without overburdening CPUs or draining power reserves.

The application required running the KamNet¹ AI model to determine a satellite's position over Earth by processing live images - demanding high processing throughput and predictable system behavior.

The Solution

The solution architecture pairs Klepsydra AI with Enclustra’s Mercury+ XU1 SoM, built around AMD’s Zynq™ UltraScale+™ MPSoC. The system runs KamNet — a convolutional neural network trained for geolocation — directly onboard. Key parts of the inference pipeline are executed in programmable logic using FPGA-based data processing units (DPUs), offloading the CPU and enabling parallel execution.

Klepsydra AI’s design prioritizes deterministic behavior, low memory variability, and real-time performance. Combined with the Mercury+ XU1’s FPGA acceleration and I/O flexibility, the result is a compact and powerful AI system ideal for deployment in orbital environments.

System Highlights with Mercury+ XU1

• Compact, High-Performance Hardware
  Measuring just 74 × 54 mm, the Mercury+ XU1 packs exceptional compute density into a space-friendly footprint.
• Powerful Compute Architecture
  Built on AMD Zynq UltraScale+ MPSoC, combining programmable logic with quad-core ARM Cortex-A53, dual-core ARM Cortex-R5, and Mali™-400 GPU.
• FPGA Acceleration for AI
  Offloads compute-intensive AI inference to FPGA fabric, delivering parallel execution, high throughput, and minimal CPU utilization.
• High-Speed Memory & I/O
  Supports up to 4 GB DDR4 SDRAM and provides extensive connectivity, including Gigabit Ethernet, USB 3.0, PCIe Gen2, and DisplayPort.
• Long-Term Reliability
  Designed with a long product lifecycle and industrial-grade robustness—making it ideal for aerospace, defense, and remote edge deployments.
• Flexible OS Compatibility
  Works seamlessly with a range of embedded operating systems, easing system integration across industries.

Klepsydra AI Efficient Software for Edge Computing

Klepsydra AI delivers deterministic, real-time performance at the edge by tightly coupling its software framework with FPGA-based hardware acceleration. At the core of this design is the use of ultra-low-latency, lock-free algorithms, complemented by a 2-dimensional threading framework that orchestrates concurrent AI tasks across both CPU and FPGA execution units.

The AI inference engine and streaming components are designed to interface directly with FPGA hardware via a dedicated connector layer. Whether deployed with or without containerization, the framework supports a range of compute architectures, including soft cores on FPGA, RISC-V, and multi-core processors. This flexibility allows Klepsydra AI to fully exploit the parallelism and deterministic timing advantages of FPGAs, making it well-suited for power- and latency-critical applications in space, defense, and industrial environments.

Applications

Apart from its application in satellite geolocation as described above, the Klepsydra AI on Enclustra’s Mercury+ XU1 SoM has a wide range of applications in the following industries:

• Aerospace: Detection and classification of objects such as ships, vehicles directly on board of the satellite or an unmanned aerial vehicle (UAV). Enabling autonomous navigation and maneuvering, fault detection, optimization of bandwidth allocation, prediction and avoidance of signal inferences, and many more.
• Defense: Enabling modern combat systems with capabilities such as autonomous vehicle operation, object identification, swarm operations, threat detection, perimeter security, and protection against cybersecurity attacks.
• Industrial Automation: Predictive maintenance of complex machines, analysis of and reaction to process data on the factory floor, detection of faulty processes and automated correction, protection against cybersecurity threats of networked systems.
• IoT: Reduction of bandwidth usage and increase of reliability by analyzing data directly on the sensor, transmitting only relevant information while filtering out noise, monitoring remote systems without excessive power consumption, and saving battery life
• Healthcare: Enabling privacy-compliant personal health devices through secure, on-device data processing, while supporting the early detection of deteriorating patient conditions in real time

Enclustra Mercury+ XU1 SoM

The Mercury+ XU1 is a compact (74 × 54 mm), high-performance SoM built around AMD’s Zynq UltraScale+ MPSoC. It offers a powerful combination of programmable logic and ARM® Cortex® processors, enabling developers to offload and accelerate AI workloads efficiently. With DDR4 support, high-speed I/O, and long product lifecycle, it's ideally suited for edge intelligence applications—on Earth or in orbit.

Results & Outlook

This proof-of-concept demonstrates that real-time AI geolocation can be performed independently in orbit using commercially available FPGA-based modules.

The collaboration between Klepsydra AI and the Enclustra Mercury+ XU1 SoM showcases a new frontier: reliable, deterministic AI inference at the edge, even in extreme environments.

As AI-driven autonomy becomes a cornerstone of aerospace, defense, and industrial systems, the Mercury+ XU1 provides the scalable, future-proof foundation needed to bring cutting-edge intelligence closer to where data is generated—whether on Earth, at sea, or in orbit.

Contact

To learn more or discuss your next Edge AI project, get in touch with us.