CheckMate: LLM-Powered Approximate Intermittent Computing

TL;DR

CheckMate targets energy constraints in batteryless IoT by shifting from minimizing checkpointing overhead to automatically generating energy-saving code approximations with controlled accuracy loss. It uses LLMs to identify and implement approximations, a robust error-resolution loop to ensure correctness, and Bayesian optimization to auto-tune knobs within a cycle-accurate intermittent computing simulator, achieving substantial power-cycle reductions. The framework is open-source and validated across six applications, five energy traces, and hardware testbeds, with results showing significant improvements over semi-automated baselines and strong real-world applicability. Formally, the trade-off is captured by the optimization metric $e_m + c_r$, where $e_m = \frac{|a_o - a_a|}{a_o}$ and $c_r = \frac{c_a}{c_o}$, balancing output fidelity and energy efficiency. Overall, CheckMate demonstrates a scalable, user-friendly path toward fully automated approximation frameworks for intermittent, batteryless IoT environments.

Abstract

Batteryless IoT systems face energy constraints exacerbated by checkpointing overhead. Approximate computing offers solutions but demands manual expertise, limiting scalability. This paper presents CheckMate, an automated framework leveraging LLMs for context-aware code approximations. CheckMate integrates validation of LLM-generated approximations to ensure correct execution and employs Bayesian optimization to fine-tune approximation parameters autonomously, eliminating the need for developer input. Tested across six IoT applications, it reduces power cycles by up to 60% with an accuracy loss of just 8%, outperforming semi-automated tools like ACCEPT in speedup and accuracy. CheckMate's results establish it as a robust, user-friendly tool and a foundational step toward automated approximation frameworks for intermittent computing.

Figures (9)

• Figure 1: CheckMate workflow.
• Figure 2: CheckMate architecture.
• Figure 3: Minimizing Optimization Metric: The purple dot indicates the point where the optimization metric achieves its minimum value, representing the balance between reduced power cycles and acceptable output error.
• Figure 4: RF Traces
• Figure 5: Optimization metrics for each LLM.Depicting performance of GPT-4o (right bar), GPT-4o Mini (middle bar), and Claude 3.5 Sonnet (left bar) across six applications. The lower values of the optimization metric ($e_m$+$c_r$) indicate better performance.