space-rha.comShort courseChapter 4 · Mechanisms and technologies

Chapter 4 · Lesson 2 of 6 · about 18 minutes

Latchup and functional interrupts

In a susceptible bulk CMOS structure, the parasitic transistors can form a regenerative path between supply and ground. The ion is long gone, but the current stays. Protection has to detect that path and interrupt it before the part overheats, the rail collapses or something else on the rail is affected.

SEL and high-current SEFI laboratory

The space-rha.com laboratory behind the guide "SEL or high-current SEFI?": fire an ion, watch the parasitic path latch, try a reset and a power cycle, and compare the diagnostic tests.

An original space-rha.com laboratory, embedded unchanged with its own controls; its onset values, gains and depths are teaching assumptions, not limits for flight hardware.

A strike can open a sustained power path.

Limiting the current is not the same as clearing the latchup.

A current limit can leave the structure above its holding condition, so the latch persists. To clear it, the rail has to fall far enough and stay down long enough for that device and that power network. A high-current event can also come from a functional state rather than a parasitic latch. Characterize the path before settling on a protection strategy.

The protection test should cover the detector, the switch, any residual supply paths, the capacitance on the rail, the restart, and repeated events in succession. Temperature, bias and how deep the ion penetrates all influence the susceptibility that is observed.

Evidence to retain. Record the rail and pin currents, how far the voltage collapsed, the trip threshold and delay, the temperature and bias cases, any residual power paths and the parametrics measured after the event. Check that the ion had range enough to reach the sensitive structure.

Mission context. A latchup on a shared instrument rail can reset channels that were not hit, and a software restart does nothing about a part that is still heating.

A current step is a symptom, not a mechanism.

A parasitic latch raises the supply current, and so can a functional state that the part gets stuck in. If every event that clears with a power cycle is called SEL, the result can be the wrong mitigation or a part rejected without cause. Keep the evidence that would let you tell the plausible causes apart.

Tell the causes apart with a controlled sequence.

Watch the function, the individual rails and the pin currents. Try the resets or commands the design permits, and then a controlled reduction of the supply followed by reinitialization. Look at the holding behavior, how long the event lasted, the bias, the temperature and the ion energy and range. No single one of these observations is conclusive across all technologies.

A reset that clears the current points toward a functional state, but it is not proof. A reset that fails to clear it may mean a true latch path, or only that the reset did not reach the part of the chip that was stuck.

Evidence to retain. Keep the rail and pin waveforms on a common time base, the functional telemetry, the reset sequence, the voltage sweep, the timing and the configuration. Where the cause could not be settled, say so and list the hypotheses that remain.

Mission context. A converter can enter a high-current control state that needs reinitialization. Protection designed only to interrupt a latchup quickly may not restore the configuration.

A part can be alive while its function is lost.

A single-event functional interrupt may stop an interface, freeze a controller, change the configuration or put the part into a mode nobody intended. Recovery may take a command, a reset, a reconfiguration or a power cycle. What the assurance case has to address is the function that was lost and whether the recovery covers it.

Exercise the control plane as well as the data plane.

Boot, standby, communication, calibration, memory management and fault handling each run on different internal circuitry. A checksum can miss a dead interface, and a watchdog can miss an output that is plausible but wrong. The workload is as much a part of the test configuration as the bias.

Classify the failures by what was observed and what it took to recover. Measure the detection latency, the data lost and whether the event repeats, using the recovery sequence the flight system will actually use.

Evidence to retain. State the software and firmware revision, the active modes, the stimuli, what the test covered, the reset domains and the recovery timing. Fault injection shows how well the system detects and recovers; the beam shows what the part does under radiation.

Mission context. A payload controller can go on answering a health poll after its acquisition engine has stopped.

Going deeper

The parasitic SCR closes a feedback loop.

The injected current produces voltage drops across the well and substrate resistances, which forward bias the parasitic junctions. The two coupled bipolar transistors then sustain the conduction, provided the regenerative gain is high enough and the external supply can deliver the current. The holding behavior, current crowding and the electrothermal evolution need more than an ideal two-transistor gain sum.

In the simplified low-injection SCR picture, αPNP + αNPN approaches one, equivalently βPNPβNPN approaches one. Real structures are distributed and the gains are transient, so a quantitative answer needs a calibrated model.

Diagnosis means comparing candidate mechanisms.

A parasitic SCR predicts a conduction path that the supply sustains and a holding condition. A logic or analog state mechanism predicts a different set of reset and mode dependencies. Self-heating, back powering and circuit states that behave like a latch can blur the two signatures. Several observations together are worth more than any one of them, and the alternatives that were not excluded should stay in the record.

A checklist of signatures helps organize the evidence, but it is not a classifier that works for every technology.

A small control-state perturbation can have global reach.

State machines, configuration registers, clocks, bias controls and shared peripherals couple a local disturbance to many outputs. SEFI is a name for a symptom. It does not identify a transistor-level mechanism, and in a commercial part whose architecture is not disclosed, working back from the symptom to the mechanism has limits.

Keep more than one hypothesis open, run the modes that would distinguish them, experiment with the reset domains and watch the pin and supply currents. Recovery on a power cycle by itself does not prove parasitic latchup.

The common mistake

The temperature response, the reset response and a voltage that looks like a holding voltage are each evidence. None of them is proof on its own.

The state-machine illustration in the laboratory cannot enumerate the hidden states of a commercial system-on-chip.

Takeaway

  • Show that the protection extinguishes the actual event and recovers from it with the real supply network in place.
  • Classify an event by the evidence and by what it took to recover before giving it a physical name.
  • A claim of recovery needs a demonstration that the event was detected, the part was isolated and the function came back.

Check yourself

A current limiter set just above the normal supply current of a CMOS part is offered as the latchup protection. What else has to be shown before it is credited?

Limiting the current is not the same as clearing the latchup. A current limit can leave the structure above its holding condition, so the latch persists and the part keeps heating; to clear it, the rail has to fall far enough and stay down long enough for that device and that power network. The demonstration should cover the detector, the switch, any residual supply paths, the capacitance on the rail, the restart and repeated events in succession, with the real supply network in place (lesson 10.3 works out what the rail capacitance alone delivers into the latched path).

In a heavy-ion run a part draws high current, a reset does not clear it and a power cycle does. The test lead logs the event as SEL. Why does the course call that premature, and what should the record hold instead?

A current step is a symptom, not a mechanism: a parasitic latch and a functional state the part is stuck in can both raise the supply current, and recovery on a power cycle by itself does not prove parasitic latchup. A reset that fails to clear the current may mean a true latch path, or only that the reset did not reach the part of the chip that was stuck; the temperature response, the reset response and a voltage that looks like a holding voltage are each evidence, and none is proof on its own. Keep the rail and pin waveforms on a common time base, the functional telemetry, the reset sequence, a controlled supply reduction followed by reinitialization, the duration, bias, temperature and ion range, and where the cause could not be settled, say so and list the hypotheses that remain. Calling every event that clears with a power cycle SEL can lead to the wrong mitigation or to a part rejected without cause.

A payload controller keeps answering its health poll throughout a heavy-ion run, and the run is reported as free of functional interrupts. What is missing?

A part can be alive while its function is lost: a controller can go on answering a health poll after its acquisition engine has stopped. A checksum can miss a dead interface and a watchdog can miss an output that is plausible but wrong, so the test has to exercise the control plane (boot, standby, communication, calibration, memory management and fault handling) as well as the data plane, and the workload is as much a part of the test configuration as the bias. Classify the failures by what was observed and what it took to recover, and measure the detection latency, the data lost and whether the event repeats, using the recovery sequence the flight system will actually use.

The latchup protection on a shared instrument rail trips quickly and restores power, and the designer regards the function as recovered. In what ways can the function still be lost?

A latchup on a shared rail can reset channels that were not hit; a high-current event can be a functional control state that needs reinitialization, which protection designed only to interrupt a latchup quickly may not restore; and a software restart does nothing about a part that is still heating. A claim of recovery needs a demonstration that the event was detected, the part was isolated and the function came back, including the restart and repeated events in succession.

A team proposes to verify the SEFI recovery of a commercial system-on-chip with fault injection alone and to skip the beam demonstration. What does each method show, and why does the course keep both?

Fault injection shows how well the system detects and recovers from the faults that were injected; the beam shows what the part does under radiation, including modes nobody anticipated. SEFI is a name for a symptom rather than a transistor-level mechanism, and in a commercial part whose architecture is not disclosed the hidden states cannot be enumerated, so the injected faults do not necessarily cover the event the beam produces. The record should state the software and firmware revision, the active modes, the stimuli, what the test covered, the reset domains and the recovery timing.

Further reading