- 學位論文
鐵電場效應電晶體記憶體陣列在記憶體增強型神經網路中最鄰近搜索之變異度分析
Variation Analysis of FeFET Memory Arrays in Nearest Neighbor Search for Memory-Augmented Neural Network
若您是本文的作者，可授權文章由華藝線上圖書館中協助推廣。
摘要
這篇論文研究了基於二氧化鉿的鐵電場效應電晶體(FeFET)在先進計算架構中的潛力和挑戰，特別是在記憶體增強型神經網路（MANN）中的應用。此架構結合了外部記憶體模組和神經網路，使模型在學習過程中能有效地儲存和檢索訊息，從而提高小樣本學習(Few-Shot Learning)的能力。進一步還可將最鄰近搜索(Nearest Neighbor Search, NNS)實現於記憶體中，通過記憶體內搜索(In-Memory-Searching)來提供一種克服記憶體牆(Memory Wall)的方法。 本論文考慮了兩種作為外部記憶體模組的FeFET陣列: (1)三元內容可定址記憶體(Ternary Content-Addressable Memory, TCAM)陣列和 (2)NOR陣列。為了評估這兩種方法的性能，本研究構建了一個根據實驗數據校準的FeFET緊湊模型(Compact Model)用於陣列級模擬。該模型包含一個變異模組，可用於評估元件變異對這兩種記憶體陣列中最鄰近搜索準確性的影響。 在FeFET TCAM陣列和NOR陣列中實現最鄰近搜索的機制不同，TCAM陣列利用匹配線(Match Line)的不同放電速率來搜索最近鄰居(在記憶體陣列中與給定資料最相似的資料)，而NOR陣列使用分壓(Voltage divider)機制來區分不同數量的不匹配位元，從而找到最近鄰居。 最後，本文進行了全面的變異分析以評估元件間變異性對最鄰近搜索和MANN準確度的影響。透過蒙特卡羅模擬顯示，由於導通電流(Ion)變異的累積，基於2FeFET的TCAM陣列易受製程變異的影響。本篇論文提出採用2FeFET再加上一電阻(2FeFET-1R)的TCAM單元可以通過抑制導通電流變異來抑制此問題。NOR陣列也提供了另一種減少由元件變異引起的最鄰近搜索準確度下降的替代方案。此外，本文通過Omniglot數據集評估了兩種陣列在MANN中的性能，結果顯示使用FeFET NOR陣列作為MANN中的外部記憶體模組，可運用於小樣本學習中，並有助於減少在推論所需的時間，且與基於軟體的計算相比表現出相近的推論準確度。
並列摘要
This thesis investigates the potential and challenges of HfO2-based ferroelectric field-effect transistors (FeFETs) in advanced computing architectures, particularly focusing on memory-augmented neural networks (MANNs). These architectures integrate an external memory module with neural networks, facilitating efficient information storage and retrieval during learning processes, thereby enhancing few-shot learning capabilities. Furthermore, implementing nearest neighbor search (NNS) within these memory modules offers a strategy to overcome the memory wall by enabling in-memory searching. Two types of FeFET arrays are examined as external memory modules: (1) ternary content-addressable memory (TCAM) arrays and (2) NOR arrays. To assess their performance, a compact FeFET model, calibrated with experimental data, is developed for array-level simulations. This model includes a variation block to evaluate the impact of device variation on NNS accuracy in both arrays. The mechanisms of NNS implementation in FeFET TCAM and NOR arrays differ. The TCAM array utilizes varying discharging rates of match lines for nearest neighbor identification, whereas the NOR array employs a voltage dividing mechanism to distin- guish the number of mismatched bits for NNS. Finally, the thesis presents a detailed variation analysis to determine the effect of device-to-device variability on NNS and MANN accuracy. Through Monte Carlo simulations, it is revealed that 2FeFET TCAM arrays are susceptible to process variation, primarily due to the accumulation of on-current variation. Employing a 2FeFET-1R TCAM cell design significantly reduces this issue by suppressing on-current variation. Alternatively, the NOR array approach demonstrates resilience against accuracy degradation in NNS caused by device variation. Additionally, the performance of both array types in MANNs is evaluated using the Omniglot dataset. The results indicate that using the FeFET NOR array as an external memory module in MANN can be applied to few-shot learning, helping to reduce the inference latency, and results in almost no decrease in inference accuracy compared to software-based computations.
參考文獻
[1] M. Lapedus, “A new memory contender?” 2018. [Online]. Available: https: //semiengineering.com/a-new-memory-contender/
[2] “Beyond CMOS - International Roadmap for Devices and Systems (IRDS),” 2021.
[3] R.Whatmore, “Piezoelectric and pyroelectric materials and their applications,”Electronic Materials: From Silicon to Organics, pp. 283–290, 1991.
[4] P. Chandra and P. B. Littlewood, “A Landau primer for ferroelectrics,” in Physics of Ferroelectrics: A Modern Perspective. Springer, 2007, pp. 69–116.
[5] S. Salahuddin and S. Datta, “Use of negative capacitance to provide voltage amplification for low power nanoscale devices,” Nano letters, vol. 8, no. 2, pp. 405–410, 2008.