ML-PCM : Machine Learning Technique for Write Optimization in Phase Change Memory (PCM)

TL;DR

This paper introduces a neural network-based method to optimize write operations in phase-change memory (PCM), significantly improving endurance, reducing latency, and lowering energy consumption for more practical data storage solutions.

Contribution

It presents a novel neural network model that predicts key PCM parameters, enhancing write efficiency and device lifespan, which is a new approach in PCM performance optimization.

Findings

NN predictions achieve 0.0073% MAPE for endurance

0.23% MAPE for total write latency

4.92% MAPE for total write energy

Abstract

As transistor-based memory technologies like dynamic random access memory (DRAM) approach their scalability limits, the need to explore alternative storage solutions becomes increasingly urgent. Phase-change memory (PCM) has gained attention as a promising option due to its scalability, fast access speeds, and zero leakage power compared to conventional memory systems. However, despite these advantages, PCM faces several challenges that impede its broader adoption, particularly its limited lifespan due to material degradation during write operations, as well as the high energy demands of these processes. For PCM to become a viable storage alternative, enhancing its endurance and reducing the energy required for write operations are essential. This paper proposes the use of a neural network (NN) model to predict critical parameters such as write latency, energy consumption, and endurance by monitoring real-time operating conditions and device characteristics. These predictions are key to improving PCM performance and identifying optimal write settings, making PCM a more practical and efficient option for data storage in applications with frequent write operations. Our approach leads to significant improvements, with NN predictions achieving a Mean Absolute Percentage Error (MAPE) of 0.0073% for endurance, 0.23% for total write latency, and 4.92% for total write energy.