As dynamic random access memory (DRAM) and other current transistor-based memories approach their scalability limits, the search for alternative storage methods becomes increasingly urgent. Phase-change memory (PCM) emerges as a promising candidate due to its scalability, fast access time, and zero leakage power compared to many existing memory technologies. However, PCM has significant drawbacks that currently hinder its viability as a replacement. PCM cells suffer from a limited lifespan because write operations degrade the physical material, and these operations consume a considerable amount of energy. For PCM to be a practical option for data storage-which involves frequent write operations-its cell endurance must be enhanced, and write energy must be reduced. In this paper, we propose SMART-WRITE, a method that integrates neural networks (NN) and reinforcement learning (RL) to dynamically optimize write energy and improve performance. The NN model monitors real-time operating conditions and device characteristics to determine optimal write parameters, while the RL model dynamically adjusts these parameters to further optimize PCM's energy consumption. By continuously adjusting PCM write parameters based on real-time system conditions, SMART-WRITE reduces write energy consumption by up to 63% and improves performance by up to 51% compared to the baseline and previous models.

翻译：随着动态随机存取存储器（DRAM）及其他当前基于晶体管的存储器逐渐接近其可扩展性极限，寻找替代存储方法的需求日益迫切。相变存储器（PCM）因其可扩展性、快速访问时间以及与许多现有存储技术相比的零泄漏功率，成为一种有前景的候选技术。然而，PCM目前存在显著缺陷，阻碍了其作为替代方案的可行性。PCM单元的使用寿命有限，因为写入操作会损耗物理材料，且这些操作消耗大量能量。要使PCM成为涉及频繁写入操作的数据存储实用选择，必须提高其单元耐久性并降低写入能量。本文提出SMART-WRITE方法，该方法集成神经网络（NN）和强化学习（RL），以动态优化写入能量并提升性能。NN模型通过监测实时运行条件和设备特性来确定最佳写入参数，而RL模型则动态调整这些参数以进一步优化PCM的能量消耗。通过基于实时系统条件持续调整PCM写入参数，与基准方法和先前模型相比，SMART-WRITE将写入能量消耗降低高达63%，性能提升高达51%。