Traditional computers are built on a fundamental assumption: power is continuous and stable. However, next-generation Internet of Things (IoT) and edge devices increasingly rely on ambient energy harvesting—such as solar, kinetic, thermal, or radio-frequency (RF) energy—to achieve true autonomy without bulky, hazardous, and short-lived batteries. As highlighted by Su et al., these ambient sources are highly stochastic, weak, and unpredictable.
A typical batteryless device operates in a cyclic fashion: it accumulates a small amount of energy in a tiny capacitor, boots up, executes a few instructions, and then suffers a brownout as the energy is depleted. In standard architectures, a power failure immediately wipes out the volatile memory (SRAM). When power returns, the device restarts from the very beginning of its program. If a computational task requires more energy than a single capacitor charge can provide, the system enters an infinite loop of restarting and crashing, failing to make any forward progress. Intermittent computing is the paradigm shift required to make execution resilient to these frequent power cycles.
Transient Computing via State Checkpointing
To break the cycle of infinite restarts, researchers developed transient computing. The core intuition is simple: if we can save the system's state to non-volatile memory (NVM) right before a crash, we can resume execution from that exact point when power returns. This effectively stretches a long computation across multiple intermittent power cycles.
As explored by Lazarescu and Balsamo, integrating this "hibernation" mechanism into standard IoT operating frameworks allows software to become resilient to sudden shutdowns. However, naive checkpointing introduces a dangerous overhead. If a system spends too much energy saving its state, it might not have enough energy left to do actual work. In worst-case scenarios, it might run out of power during the checkpointing process itself, leading to state corruption.
Traditional Volatile Execution:
[Charge] -> [Execute Task...] -> [Power Loss / Wipe SRAM] -> [Restart from Scratch] (Livelock)
Transient Execution with Checkpointing:
[Charge] -> [Execute Part 1] -> [Save Checkpoint to NVM] -> [Power Loss] -> [Recover] -> [Execute Part 2]
To minimize this overhead, Zhang et al. proposed tracking data flow through hardware using low-cost binary counters. By understanding exactly which pieces of data are essential for the next stage of computation, the system can selectively save only the critical intermediate data rather than the entire memory state. This reduces the time and energy spent on checkpointing to near zero, maximizing the energy left for forward progress.
Power-Neutral and Energy-Driven Execution
While transient computing focuses on surviving crashes, an alternative approach aims to avoid them altogether by matching the system's consumption to the environment's generation. This is known as power-neutral computing or energy-driven computing.
As argued by Al-Hashimi et al., instead of using large hardware buffers to mimic a steady battery, systems should be designed to run directly on highly variable power. To achieve this, Balsamo et al. demonstrated that systems can use adaptive frequency and voltage scaling to dynamically throttle or boost performance in real time based on incoming energy. If the sun goes behind a cloud, the processor slows down, lowering its power draw to match the weaker solar input, thereby maintaining continuous operation without crashing.
Taking this elasticity to the extreme, Mileiko et al. redesigned hardware components using temporal encoding (Pulse Width Modulation) rather than binary voltages. Because information is encoded in the timing of pulses rather than absolute voltage levels, the underlying circuits remain functional and accurate even when the supply voltage fluctuates wildly, making the hardware inherently "power-elastic."
Scaling to Complex Workloads and Distributed Learning
As edge devices are increasingly tasked with running artificial intelligence, intermittent computing must scale to handle complex workloads like deep learning inference and distributed training.
For local inference, Wu et al. introduced multi-exit neural networks optimized for intermittent execution. Deep neural networks normally require massive, uninterrupted computational blocks. By embedding "early exits" throughout the network and compressing layers non-uniformly, the system can output a lower-accuracy prediction if power is running low, or proceed to deeper, more accurate layers if a sustained burst of energy is available.
Intermittent Inference (Multi-Exit NN):
[Input] -> [Layer 1] -> [Layer 2] -> [Early Exit 1] ---> (Low Energy? Output Result Now)
|
[Layer 3] -> [Final Exit] ---> (High Energy? Output Full Accuracy)
When scaling to distributed environments, such as federated learning, the challenge shifts from individual node crashes to network-wide coordination. Traditional federated learning assumes that client devices are always powered and ready to synchronize their gradients. However, as Yener and Guler and Shen et al. point out, the stochastic nature of energy harvesting means edge clients appear and disappear unpredictably.
To train unbiased models under these constraints, the authors developed scheduling protocols that scale gradients based on local energy arrival statistics and allow asynchronous updates. This ensures that the global model converges accurately even when individual participants are constantly losing power and dropping out of the training loop.