採用四位元電阻式閘極鰭式電晶體的仿生計算平台

Abstract

本研究首先利用電阻式記憶體 (RRAM) 和鰭式電晶體 (FinFET)，建構出電阻式閘極鰭式電晶體 (RG-FinFET) 的架構。RG-FinFET 的記憶體結構是將一個電阻式記憶體串接在高介電係數鰭式電晶體的閘極，藉由調變 RRAM 的阻態以切換 FinFET，從而通過 RG-FinFET的汲極電流來區分內部儲存的狀態。在 RRAM 元件特性上，因有 HfON 作為介電層，具有無須Forming的特性，不需要太大的電流即可操作，能大幅降低功耗，且有良好的穩定性。與FinFET整合成電阻式記憶體之後的 RG-FinFET 可以在超低電流下操作。操作速度方面，最快能在 10 奈秒內完成 SET，在 4 奈秒內完成 RESET。除此之外，此元件能利用較小的脈衝電壓來達成對稱的漸變式操作，此記憶體最大與最小的讀出電流比有 106 倍，可明確區分成16個狀態 (4-bit-per-cell)。在可靠度的測試中，16個狀態都可以操作超過 105 次，且預估在 138°C 的高溫烘烤下可以維持十年。RG FinFET 在記憶體陣列中的特性也加以評估，內部閘極電介質可以有效防止潛行路徑 (sneak path)，避免資訊從鄰近單元流失。在電性干擾的分析中，不論是編程或讀取的干擾，都具有很好的抗干擾能力。其次，本研究將 RG-FinFET 作為基本元件，建構出全由電阻所組成的神經仿生計算平台，此平台包含權重儲存、激發函數電路以及類比數位轉換器，皆是以 RG-FinFET 為核心完成。本研究提出的阻式整流線性單位函式 (rReLU) 的功能相當理想，電阻式類比數位轉換器 (rADC) 也有很好的轉換結果，向全電阻式仿生計算平台 (FRNC) 的發展更近了一步。

We complete the architecture of the resistive-gate FinFET (RG-FinFET) with an integration of Resistance Random Access Memory (RRAM) and a FinFET. The RG-FinFET is a structure which connects an RRAM and the gate of high-k FinFET in series. By modulating the resistance of the RRAM, we can change the states of the FinFET. In other words, we can distinguish the 0 and 1 states and store the states from the varying drain currents of the RG-FinFET.In terms of the RRAM characteristics, because HfON is used as a dielectric layer, it has a characteristic of forming-free. It requires less current to operate, which can greatly reduce power consumption and has good stability. As a unit cell of RG-FinFET, it can operate at ultra-low current. In terms of operation speed, the SET operation can be completed in 10 nanoseconds, and the RESET can be completed in 4 nanoseconds. In addition, this device can achieve symmetrical gradual operation with a small pulse voltage. The maximum and minimum read current ratio of the memory is 10^6 times, which can be clearly distinguished into 16 states (4-bit-per-cell). For the reliability test, 16 states can be operated more than 10^5 times, and it is estimated that it can last for ten years under the high temperature baking of 138°C. The characteristics of RG FinFETs in memory arrays are also evaluated, and the internal gate dielectrics can effectively prevent sneak path, preventing information loss from adjacent cells. For the analysis of disturb, whether it is programming or reading, it has good disturb immune ability.Furthermore, in this work, we used RG-FinFET as the basic component to construct a neuromorphic computing platform composed of resistors. This platform includes weight storage, activation function circuit, and analog-to-digital converter, all of which are constructed by the using of RG-FinFET. The resistive rectification linear unit function (rReLU) proposed in this study works quite well. The resistive analog-to-digital converter (rADC) also has good conversion results. As a result, we are one step closer to a fully resistive neuromorphic computing platform (FRNC).