Traditional embedded systems are designed with a fundamental assumption: power is a constant. When the plug is pulled or the battery dies, the system crashes, volatile memory is wiped, and the program restarts from scratch upon power restoration. As the Internet of Things scales toward billions of devices, the maintenance burden of batteries has become unsustainable. This has led to a paradigm shift toward energy-driven computing, where devices harvest ambient energy—such as solar, vibration, or thermal gradients—and treat power unpredictability as a first-class citizen.
The Problem of the Reboot Loop
In an energy-harvesting environment, power is often available only in short, unpredictable bursts. A device might wake up, begin a calculation, and lose power before finishing. Without a specialized architecture, the device reboots when energy returns, starts the same calculation, and fails again at the same point.
As noted in Energy-driven computing by Sivert T. Sliper, Alex S. Weddell, and colleagues, a critical failure mode occurs when a task requires more energy than the system's storage buffer can provide. This results in a perpetual reboot cycle where no forward progress is ever made. To solve this, the system must be able to "pause" and "resume" across power cycles.
Reactive Checkpointing: Just-in-Time Persistence
The core mechanism for surviving intermittency is state checkpointing—saving the CPU registers and RAM to non-volatile memory (NVM) before power vanishes. Early approaches used periodic checkpointing, but this is inefficient; saving state consumes energy that could have been used for computation.
A more elegant solution is reactive state checkpointing, exemplified by the Hibernus system described in Energy-driven computing for energy-harvesting embedded systems. Instead of saving state on a timer, the system uses hardware-level voltage monitoring. When the supply voltage drops below a critical threshold, a hardware interrupt triggers a single, high-speed save of the entire system state to NVM.
This approach works because it maximizes the time spent on "useful" instructions. By delaying the save until the last possible millisecond, the system ensures that energy is only spent on persistence when a failure is truly imminent. Theodoros D. Verykios and Domenico Balsamo further optimized this in Selective policies for efficient state retention in transiently-powered systems by demonstrating that saving only the "allocated" or used portions of memory can reduce energy overhead by up to 85%.
Beyond the CPU: The Peripheral Challenge
While restoring a CPU's registers is straightforward, restoring the state of the entire system is not. Modern IoT devices rely on external sensors and radio modules that have their own internal states. Domenico Balsamo points out in Transient and power-neutral computing: a paradigm shift for embedded systems? that re-establishing a radio connection or re-calibrating a sensor after a power cycle is often more energy-intensive than the computation itself.
To address this, researchers have developed "comprehensive state restoration" techniques. As detailed in ARM mbed support for transient computing in energy harvesting IoT systems, this involves capturing and reapplying the configuration of peripheral registers. This ensures that when the device "wakes up," the radio doesn't just reboot—it resumes its protocol state as if the power never flickered.
Hardware-Native Intermittency
The ultimate evolution of this field moves away from software-managed checkpoints toward hardware that is inherently non-volatile. Salonik Resch and colleagues proposed RODENT in Towards Homomorphic Inference Beyond the Edge, which utilizes non-volatile processing-in-memory (PIM).
By using spintronic or other non-volatile state elements directly in the computational logic, the system achieves hardware-level idempotency. In this model, the computational state is maintained in the hardware itself. There is no "save" or "restore" phase; the system simply stops when power is gone and continues from the exact same gate-level state when power returns. This eliminates the energy waste associated with software checkpointing entirely.
Applications in Extreme Environments
These mechanisms enable complex tasks in environments where batteries cannot be replaced. For instance, Yushan Huang demonstrated in Towards Machine Learning and Inference for Resource-constrained MCUs that battery-free microcontrollers can perform deep-sea fish recognition using harvested energy. By partitioning neural network inference into resumable segments, these devices can complete high-accuracy machine learning tasks across dozens of power interruptions, transforming tiny, erratic bursts of environmental energy into sustained computational progress.