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.

What problem does this paper attempt to address?

### What problems does this paper attempt to solve? This paper aims to solve the problems of energy limitation and excessive checkpoint overhead in battery - free Internet of Things (IoT) systems. Specifically: 1. **Energy Constraints**: Battery - free IoT devices rely on ambient energy harvesting, which leads to instability and unpredictability of energy supply, making frequent and unpredictable power failures a common phenomenon. Therefore, achieving efficient energy utilization in such an environment is crucial. 2. **Checkpoint Overhead**: Traditional intermittent computing methods deal with power failures by frequently saving and restoring program states, but this introduces a large amount of overhead. Although previous studies have attempted to reduce these overheads by means such as optimizing checkpoint placement and virtualizing memory, the room for further optimization is already very limited. 3. **Limitations of Manual Tuning**: Existing approximate computing techniques require a great deal of manual tuning, which limits their scalability and applicability in different applications. Developers must find a balance between energy savings and computational accuracy, and this process is both time - consuming and error - prone. ### Solutions To solve the above problems, the paper proposes the CheckMate framework, an automated framework based on large - language models (LLMs) for context - aware code approximation. The main features of CheckMate include: - **Automated Approximation**: Automatically generate approximate code through LLMs and combine Bayesian optimization to automatically adjust approximation parameters without the need for developer manual intervention. - **Verification Mechanism**: Integrate a verification mechanism to ensure that the approximate code generated by LLMs can be correctly compiled and run, avoiding performance degradation due to incorrect code. - **Optimization Objectives**: Minimize the optimization metric, reducing the power cycle while keeping the output error within an acceptable range. ### Specific Contributions - **Significant Energy Savings**: Tested in six different IoT applications, CheckMate reduces the power cycle by 15% to 60% while keeping the error rate between 6% and 25%. - **Outperforms Semi - automated Tools**: Compared with existing semi - automated tools such as ACCEPT, CheckMate performs better in terms of speed and accuracy and does not require expert intervention. - **User - friendly**: Through a simplified one - click process, it significantly reduces the time and error rate required for applying approximation, enabling developers to optimize battery - free IoT applications more efficiently. ### Summary By introducing intelligent automation and LLMs technology, CheckMate has successfully solved the problems of high energy consumption and complex manual tuning in battery - free IoT systems, providing a new solution for future intermittent computing.

LLM-Powered Approximate Intermittent Computing

Approxify: Automating Energy-Accuracy Trade-offs in Batteryless IoT Devices

LMM: A Fixed-Point Linear Mapping Based Approximate Multiplier for IoT

Optimally Approximated and Unbiased Floating-Point Multiplier with Runtime Configurability

An Efficient and Flexible Learning Framework for Dynamic Power and Thermal Co-Management

PAM: A Piecewise-Linearly-Approximated Floating-Point Multiplier with Unbiasedness and Configurability

Approximate Computing for Low Power and Security in the Internet of Things

Intermittent Learning: On-Device Machine Learning on Intermittently Powered System

Memory-Aware Partitioning of Machine Learning Applications for Optimal Energy Use in Batteryless Systems

Optimal Checkpointing for Secure Intermittently-Powered IoT Devices

Exploiting Significance of Computations for Energy-Constrained Approximate Computing

Designing of Energy-Efficient Approximate Multiplier Circuit for Processing Unit of IoT Devices

Energy-Latency Computation Offloading and Approximate Computing in Mobile-Edge Computing Networks

Approx-IMC: A general-purpose approximate digital in-memory computing framework based on STT-MRAM

Delay-Optimal Computation Offloading for Computation-Constrained Mobile Edge Networks.

Incidental Computing on IoT Nonvolatile Processors

ADEPOS: A Novel Approximate Computing Framework for Anomaly Detection Systems and its Implementation in 65nm CMOS

A Study on Approximate Computing for Non-volatile Memory-Based Memory Systems

IAA: Incidental Approximate Architectures for Extremely Energy-Constrained Energy Harvesting Scenarios Using IoT Nonvolatile Processors.

Memristor-based Approximated Computation.

LLM-based Frameworks for Power Engineering from Routine to Novel Tasks