AlphaRA: An AlphaZero based approach to Redundancy Analysis

Manufacturing flaws in memory devices give rise to faulty cells rendering the chips unusable and consequently reducing the wafer yield. To repair faulty memory cells, redundancies are included in the form of spare rows and columns in the memory. Redundancy Analysis is the process of mapping these spare rows and columns to repair faulty lines in the chip. However, Redundancy Analysis is an NP-complete problem, making it difficult to find a trade-off between repair rate and runtime, especially for large chip sizes. In this paper, we introduce AlphaRA, a first-of-its-kind memory repair algorithm based on the Reinforcement Learning algorithm AlphaZero. We explicate AlphaRA as a single agent problem that learns the strategies of Redundancy Analysis through self-play. Starting tabula rasa, AlphaRA achieves an average normalized repair rate of 99.8% on 16×16 chips with only 32 MCTS simulations. It outperforms the next best heuristic algorithm by 5.42% while utilizing 0.29% lesser spares, making it a suitable Redundancy Analysis algorithm for mass production of memory devices.