IN Brief:
- Mercury’s space architecture combines several processor types within open modular hardware formats.
- Intended workloads include sensor fusion, electronic warfare, navigation, inference, and autonomous fault detection.
- Onboard processing can reduce downlink demand, although radiation, power, heat, and software assurance remain limiting factors.
Mercury Systems has outlined a heterogeneous processing architecture for radiation-tolerant space systems, combining CPUs, GPUs, FPGAs, and neural processors across onboard sensing, communications, navigation, and autonomous operations.
Processing information before it is transmitted to the ground allows sensor data to be filtered, classified, compressed, or fused locally. Raw imagery, spectral data, and communications traffic need not compete continuously for limited downlink capacity when useful information can be identified aboard the spacecraft.
Potential workloads include intelligence, surveillance and reconnaissance, signals intelligence, electronic warfare, spectrum monitoring, jam detection, navigation, and equipment-health analysis. Each imposes a different balance of arithmetic throughput, memory bandwidth, latency, determinism, power consumption, and fault tolerance.
SpaceVPX, SOSA, and OpenVPX standards provide modular hardware structures that can be configured around different missions. Open interfaces also reduce dependence on a proprietary board format when processing, communications, or sensor functions are upgraded.
Mercury’s SCFE6933 is a radiation-tolerant 6U SpaceVPX board based on an AMD Versal AI Core adaptive system-on-chip. Programmable logic, processor resources, and AI engines support beamforming, software-defined radio, image processing, machine-learning inference, and other high-rate signal functions.
Within that device, programmable logic can handle deterministic interfaces and data paths while processor cores coordinate the application and AI engines execute supported inference workloads. Keeping those resources in one component reduces some board-level data movement and allows processing to be partitioned around mission constraints.
Onboard intelligence changes the communications balance
Earth-observation, signals, and defence payloads can generate more information than a spacecraft can transmit continuously. Sending every image, spectrum capture, or sensor stream to the ground introduces delay and consumes bandwidth shared with telemetry, command traffic, and other mission data.
Local processing can select cloud-free imagery, identify a change on the ground, isolate an unusual emitter, or compress a signal into its useful characteristics. Downlink capacity can then be directed towards material already judged relevant rather than a complete unprocessed dataset.
That autonomy increases the consequence of a classification error. A false negative may cause valuable information to be discarded before an operator can examine it, while a false positive can consume limited communications and storage resources.
Charged particles introduce faults rarely encountered by terrestrial AI systems, including altered memory contents, disrupted logic, and cumulative damage to semiconductor structures. Error correction, redundancy, configuration scrubbing, fault detection, and recovery mechanisms must surround the processing architecture.
Thermal control is similarly restrictive because high-performance accelerators cannot rely on conventional convective cooling in vacuum. Available power may also change with orbit, solar-array orientation, battery condition, and the demands of other spacecraft subsystems.
Model updates create a further assurance problem. Revised inference software must be authenticated, stored safely, activated without losing the previous configuration, and monitored after deployment so that unexpected behaviour can be detected and reversed.
Long service lives make semiconductor planning inseparable from architecture, as examined in work on defence-electronics obsolescence and qualification. Devices selected early in a programme may need to remain available, replaceable, or reproducible long after their commercial counterparts have left production.
At the RF end of the signal chain, silicon-germanium beamforming ICs are increasing integration within defence radar. Onboard heterogeneous processing extends the same direction into classification, fusion, and decision support after signals have been acquired.
Matching each workload to a suitable compute engine can improve performance per watt compared with forcing every operation through one architecture. Adoption will nevertheless depend on radiation evidence, thermal design, software tools, export controls, long-term availability, and predictable behaviour when faults occur.
