Edge computing in space

The case for processing data in orbit rests on one number: satellites generate far more data than they can transmit. Everything else follows from closing that gap.

The gap

The mismatch is more than an order of magnitude, and the conventional responses, duty cycling the sensor and compressing aggressively, both discard information that was already collected. See Earth observation.

Edge processing changes the calculation entirely. If what the customer wants is a ship position, a change alert or a cloud-free scene selection, then the useful output is kilobytes and the gigabytes never need to leave orbit.

The latency column is often worth more than the volume column. Maritime surveillance, disaster response and defence tasking all lose value with age, and an answer in minutes over a low-rate link can be worth more than an image in hours.

What it costs

Onboard processing is not free, and the costs land on the two most constrained budgets on a small spacecraft.

Power. An inference accelerator drawing 10 to 30 W is a substantial fraction of a smallsat’s orbit-average power. From Power systems, that is roughly 30 to 90 m² of array-equivalent generation at typical efficiencies, though duty cycling mitigates it since inference runs only while imaging.

Thermal. Every watt consumed becomes heat that must be radiated. From Thermal control, about 2.6 m² of radiator per kilowatt at room temperature. At tens of watts this is manageable; it is the constraint that bounds scaling, as discussed in Orbital data centers.

Mass and volume. Accelerator, memory, cooling and shielding all compete with payload.

The trade closes when the downlink saved exceeds the resources consumed, which is straightforward for a high-rate sensor and marginal for a low-rate one.

Radiation

Commercial accelerators are not radiation hardened, and a rad-hard processor is three orders of magnitude too slow to run inference usefully. The workable approach is commercial silicon with fault tolerance built around it.

Two properties of inference workloads make this easier than it sounds.

Inference is stateless and repeatable. A corrupted result can simply be recomputed, unlike a control loop where a corrupted output has already acted.

Neural networks are somewhat tolerant of bit errors. A flipped bit in a weight usually perturbs the output slightly rather than catastrophically, particularly in quantised networks. This is not a substitute for ECC but it raises the threshold at which upsets matter.

The consequence is that an accelerator that would be unacceptable for spacecraft control is acceptable for payload processing, which is exactly the division described in Autonomous spacecraft.

What to run onboard

The branch at the bottom is the one that matters operationally. Onboard decisions are irreversible. A cloud classifier that wrongly rejects a scene has destroyed an observation nobody can retake, and no amount of ground processing recovers it.

The safer architecture prioritises rather than deletes wherever storage allows, so a wrong triage decision costs latency rather than data.

The pattern is consistent: run onboard whatever reduces data volume or reduces latency for a time-critical answer, and leave everything else on the ground where computation is cheap and reversible.

Flight demonstrations

Phi-Sat-1, 2020

ESA CubeSat carrying a commercial vision accelerator running a neural network to detect and discard clouded images before downlink. The first flight of a deep neural network for Earth observation onboard.

HPE Spaceborne Computer on the ISS

Demonstrated that commercial servers can operate in orbit with software-based error mitigation rather than rad-hard hardware.

Commercial constellations

Several operators now fly onboard processing for detection and triage as an operational capability rather than a demonstration.

Deep space

Far more constrained. Interplanetary missions face harsher radiation and cannot use commercial accelerators, so onboard analysis remains classical.

Phi-Sat-1 is the reference result because it demonstrated the whole argument in one experiment: a commercial accelerator, a learned model, a real operational benefit in downlink saved, and an acceptable consequence when the model is wrong.

The limits

Edge computing in space is a real and deployed capability with clear boundaries.

It works for data already generated in orbit where the output is much smaller than the input. It does not make orbit a good place to run general computation: terrestrial compute is vastly cheaper per operation and always will be, and the thermal constraint bounds how far the concept scales.

The honest framing is that this is a bandwidth optimisation, not a computing platform. That is a valuable thing to be, and it is a narrower claim than the one usually made.

References

• Giuffrida, G. et al. “The Phi-Sat-1 Mission: The First On-Board Deep Neural Network Demonstrator for Satellite Earth Observation”, IEEE Transactions on Geoscience and Remote Sensing, 60:1, 2022.
• Denby, B. and Lucia, B. “Orbital Edge Computing: Nanosatellite Constellations as a New Class of Computer System”, ASPLOS 2020.
• Furano, G. et al. “Towards the Use of Artificial Intelligence on the Edge in Space Systems”, IEEE Aerospace and Electronic Systems Magazine, 35:44, 2020.
• Kothari, V. et al. “The Final Frontier: Deep Learning in Space”, HotMobile 2020.
• Bradley, A. and Fernandez, M. “Spaceborne Computer: Results from the ISS experiment”, HPE technical reports.