Before You Sign: Interrogating Fault Containment in Airborne Edge Compute

Two processing lanes on one board don't make a module fault tolerant. What decides it is whether a fault in one lane, its power rail, its clock or its configuration memory can be stopped before it corrupts the output the flight stack trusts. A procurement engineer with little lab time can't measure that directly, so each question to the vendor has to work as a probe that shows where the containment boundary really sits. The questions below are ordered by how much risk each one exposes when the vendor's answer is vague.

Where does one lane stop and the other begin?

Start by asking for the shared resource list, not the redundancy diagram. Datasheets for duplex sensor fusion modules often show two independent compute lanes, but those lanes may still draw on one point of load regulator, one reference oscillator, one boot flash or one backplane connector. Each shared element is a common-mode path, and the failure that matters most is exactly that: a brownout on a shared rail resets both lanes in the same microsecond, and the comparator that was supposed to catch the disagreement goes down with them. Ask the vendor to name every component whose failure takes out both lanes, and ask how each one is covered by monitoring, by dissimilarity or by an argument in the safety assessment. If the list comes back empty, they haven't done the analysis. They haven't removed the risk either.

What happens between the fault and the safe state

The number to pin down is the fault-to-safe-state latency at the module output, and for a multi-sensor perception pipeline feeding an avoidance function, a target such as 50 ms from first erroneous output to either a flagged invalid frame or a clean handover is a reasonable starting point to negotiate. What matters in practice is how that time is spent. Detection by cross-lane comparison is fast but only catches divergence, so a silent error both lanes compute the same way, like a shared calibration table that has gone bad, gets through untouched. Detection by watchdog or by plausibility checks against an independent sensor catches more, but adds latency and false alarms. Ask the vendor which mechanism owns each fault class, and ask for the measured latency distribution, not the design target. As an estimate, expect the tail to be several times the median once the bus is busy with heavy lidar point clouds.

Configuration memory and radiation

SRAM-based FPGAs hold their logic in configuration cells that can flip from a single event upset at altitude, so the question isn't whether upsets happen but how fast the device finds and repairs them. Scrubbing rate, error correction on configuration frames and whether the design uses triple modular redundancy in its critical paths all decide whether a flipped bit becomes a transient glitch or a lane that keeps quietly producing plausible but wrong detections. Flash-based or antifuse parts avoid most of this but give up density and reprogrammability, which matters if you expect to update perception models in the field. Make the vendor say which fabric technology carries the safety function and what upset rate they assumed for your operating altitude band.

Evidence you can check without a lab

The airborne design assurance guidance this market works under, DO-254 for complex electronic hardware and the matching environmental qualification standards, asks for verification evidence, and that evidence is your proxy for the lab work you can't do. Ask for fault injection reports done at the lane boundary: forced rail sags, clock stops, corrupted configuration frames, a stuck output on a single sensor. Ask too for the coverage figure each campaign produced. A vendor working to the higher assurance levels will have this material or know exactly why some of it doesn't exist yet. The decision rule is simple. If the vendor can't show injected-fault results for every shared resource on their own list, score the redundancy as nominal and price the module as single-lane, because that is how it will behave on the day it matters. That still leaves a trade-off to settle with your systems team: a true dual-lane module with separate power and timing costs mass and power, and as an estimate a well-isolated duplex design consumes noticeably more than half again the power of its simplex counterpart, so the next decision is whether that budget is better spent on a second, dissimilar module elsewhere in the airframe.

Is cross-lane comparison enough to claim the module fails safe?

No. Comparison only catches faults that make the lanes disagree. Errors from shared inputs, shared calibration data or a common design flaw produce the same wrong answer in both lanes, so you need an independent plausibility check or a dissimilar channel to cover them.

Back to AirspaceCompute · Domains

← Back to AirspaceCompute
Engineering & Architecture Brief

Before You Sign: Interrogating Fault Containment in Airborne Edge Compute

Two processing lanes on one board don't make a module fault tolerant. What decides it is whether a fault in one lane, its power rail, its clock or its configuration memory can be stopped before it corrupts the output the flight stack trusts. A procurement engineer with little lab time can't measure that directly, so each question to the vendor has to work as a probe that shows where the containment boundary really sits. The questions below are ordered by how much risk each one exposes when the vendor's answer is vague.

Where does one lane stop and the other begin?

Start by asking for the shared resource list, not the redundancy diagram. Datasheets for duplex sensor fusion modules often show two independent compute lanes, but those lanes may still draw on one point of load regulator, one reference oscillator, one boot flash or one backplane connector. Each shared element is a common-mode path, and the failure that matters most is exactly that: a brownout on a shared rail resets both lanes in the same microsecond, and the comparator that was supposed to catch the disagreement goes down with them. Ask the vendor to name every component whose failure takes out both lanes, and ask how each one is covered by monitoring, by dissimilarity or by an argument in the safety assessment. If the list comes back empty, they haven't done the analysis. They haven't removed the risk either.

What happens between the fault and the safe state

The number to pin down is the fault-to-safe-state latency at the module output, and for a multi-sensor perception pipeline feeding an avoidance function, a target such as 50 ms from first erroneous output to either a flagged invalid frame or a clean handover is a reasonable starting point to negotiate. What matters in practice is how that time is spent. Detection by cross-lane comparison is fast but only catches divergence, so a silent error both lanes compute the same way, like a shared calibration table that has gone bad, gets through untouched. Detection by watchdog or by plausibility checks against an independent sensor catches more, but adds latency and false alarms. Ask the vendor which mechanism owns each fault class, and ask for the measured latency distribution, not the design target. As an estimate, expect the tail to be several times the median once the bus is busy with heavy lidar point clouds.

Configuration memory and radiation

SRAM-based FPGAs hold their logic in configuration cells that can flip from a single event upset at altitude, so the question isn't whether upsets happen but how fast the device finds and repairs them. Scrubbing rate, error correction on configuration frames and whether the design uses triple modular redundancy in its critical paths all decide whether a flipped bit becomes a transient glitch or a lane that keeps quietly producing plausible but wrong detections. Flash-based or antifuse parts avoid most of this but give up density and reprogrammability, which matters if you expect to update perception models in the field. Make the vendor say which fabric technology carries the safety function and what upset rate they assumed for your operating altitude band.

Evidence you can check without a lab

The airborne design assurance guidance this market works under, DO-254 for complex electronic hardware and the matching environmental qualification standards, asks for verification evidence, and that evidence is your proxy for the lab work you can't do. Ask for fault injection reports done at the lane boundary: forced rail sags, clock stops, corrupted configuration frames, a stuck output on a single sensor. Ask too for the coverage figure each campaign produced. A vendor working to the higher assurance levels will have this material or know exactly why some of it doesn't exist yet. The decision rule is simple. If the vendor can't show injected-fault results for every shared resource on their own list, score the redundancy as nominal and price the module as single-lane, because that is how it will behave on the day it matters. That still leaves a trade-off to settle with your systems team: a true dual-lane module with separate power and timing costs mass and power, and as an estimate a well-isolated duplex design consumes noticeably more than half again the power of its simplex counterpart, so the next decision is whether that budget is better spent on a second, dissimilar module elsewhere in the airframe.

Technical FAQ: Is cross-lane comparison enough to claim the module fails safe?

No. Comparison only catches faults that make the lanes disagree. Errors from shared inputs, shared calibration data or a common design flaw produce the same wrong answer in both lanes, so you need an independent plausibility check or a dissimilar channel to cover them.

Related Architecture Specifications (AIRSPACEGATEWAY): flexiblesky.com · uavorchestration.com · agenticaav.com

More Technical Articles on airspacecompute.com

Benchmarking Teraflop Edge Compute Performance in Unpressurized High-Altitude Cargo UAVs →Mitigating Multi-Core Contention in Avionics Mission Computers Using Hardware Memory Regulators →