Comparative Study on Device Type Configurations of 2T0C DRAM for Compute-in-Memory Applications

Abstract

In this article, we systematically analyze the electrical characteristics of 2T0C DRAM unit cells for compute-in-memory (CIM) applications, focusing on the on/off current ratio, coupling effects, and retention time, with respect to the NN (NMOS-NMOS), NP (NMOS-PMOS), PN (PMOS-NMOS), and PP (PMOS-PMOS) device types. We designed 65 nm CMOS-based 2T0C DRAM unit cells using Sentaurus 3D TCAD simulation, while keeping the physical parameters and doping concentrations of the NMOS and PMOS transistors identical to impartially compare their electrical characteristics. As a result, the NP- and PN-type 2T0C DRAM cells exhibited much shorter retention times for data ‘0’ (<0.55 ns) and smaller on/off current ratios (<4.36) than the NN- and PP-type cells, clearly indicating that the NP- and PN-type cells are not favorable for CIM applications. Furthermore, we investigated the cause of the storage node voltage fluctuation immediately after a read pulse using the equivalent RC models of the NN- and PP-type cells. The origin of this fluctuation is attributed to three factors: the coupling effects between the storage node and the read transistor, the drain current in the read transistor, and the subthreshold leakage current in the write transistor.

1. Introduction

Recent advancements in artificial intelligence (AI) have enabled the efficient processing of complex tasks, such as natural language processing and image classification, so the demand for AI is rapidly increasing. In addition, the number of parameters and computations in recent AI models, including large language models (LLMs), have increased exponentially to improve computational performance [1]. However, executing AI models on the conventional von Neumann architecture results in high levels of power consumption and significant latency due to the memory bottleneck caused by the data movement between the processor and memory. To address these challenges, compute-in-memory (CIM) architectures have been actively studied as an alternative to the von Neumann architecture [2,3,4,5,6]. By enabling data storage and parallel computation within the memory array, CIM architectures minimize the data movement, thereby reducing power consumption and enhancing computing performance for AI workloads.

2T DRAM has been widely studied for CIM applications because of its advantages including its high-speed, low-power, non-destructive read operations, and CMOS compatibility [7,8,9,10,11,12,13,14,15]. However, 2T DRAM suffers from a short retention time due to its small storage node capacitance and relatively large leakage current. To improve the retention time, vertical transistor on gate (VTG) 2T DRAM structures have been proposed [9,11], which enhance retention time and minimize unit cell area compared to planar 2T0C DRAM structure. Nevertheless, since the retention time of silicon channel-based 2T DRAM remains short, further studies are required to improve its retention time. Additionally, oxide-semiconductor field-effect transistor (OSFET)-based 2T DRAM has been proposed [16,17,18,19,20,21] offering a long retention time due to its extremely low leakage current. However, recent studies on OSFET-based 2T DRAM have reported that the endurance of OSFETs is limited to approximately 10¹² cycles [21], leading to variations in the retention and transfer characteristics through iterative write and read operations. Therefore, further investigation of OSFETs is required to enhance their endurance characteristics for integration into 2T DRAM.

Moreover, no study has systematically clarified which 2T DRAM device types—NN (NMOS-NMOS), NP (NMOS-PMOS), PN (PMOS-NMOS), or PP (PMOS-PMOS)—are favorable for CIM applications. Most existing studies on silicon channel-based 2T DRAM have individually focused on the NN- [7,8,9,10,11], NP- [12], PN- [13,14,22,23], and PP-type [15,24,25] cells. In this work, we designed 65 nm CMOS-based 2T0C DRAM unit cells with the NN, NP, PN, and PP device types using Sentaurus 3D TCAD simulation and investigated their electrical characteristics, including the on/off current ratio, coupling effects, and retention time. Additionally, we identified the cause of the storage node voltage (V_SN) fluctuation immediately after a read pulse using the equivalent RC models of the NN- and PP-type cells. Since the storage node capacitance of the silicon channel-based 2T0C DRAM unit cell is very small and its subthreshold leakage current is relatively large, the retention time is easily affected by the V_SN fluctuation and retention degradation. Therefore, it is essential to clarify the cause of this fluctuation. We conducted an in-depth investigation of its origin and identified three factors: (i) the coupling effects between the storage node (SN) and the channel of the read transistor (read Tr.), read wordline (RWL), and read bitline (RBL); (ii) the drain current in the read Tr.; and (iii) the subthreshold leakage current in the write transistor (write Tr.).

2. Design of 2T0C DRAM Unit Cell Structure

Figure 1 illustrates the schematic diagrams of 2T0C DRAM unit cells with the NN, NP, PN, and PP device types, including SN and parasitic capacitive components between SN and the write wordline (C_SN-WWL), RWL (C_SN-RWL), and RBL (C_SN-RBL). In the 2T0C DRAM unit cells, the drain of the write Tr. is connected to the gate of the read Tr. Charges (data) are stored in the parasitic capacitive components of SN by activating the write Tr., and V_SN modulates the read current (I_READ) in the read Tr. during the read operation. As shown in Figure 2, we designed 65 nm NMOS and PMOS transistors based on the physical parameters in

Table 1. Physical parameter of 65 nm planar transistor.

Parameter	Value
Gate length	65 nm
Gate width	100 nm
Channel length	45 nm
Gate oxide thickness	1.7 nm
Source/drain junction depth	25 nm

using Sentaurus 3D TCAD simulation. To impartially compare 2T0C DRAM unit cells, the physical parameters and doping concentrations of the NMOS and PMOS transistors were kept identical, except for the dopant species. Figure 3 shows the transfer characteristics of the NMOS and PMOS transistors with the threshold voltages (Vth) of 0.29 V and −0.30 V, respectively. Subsequently, 65 nm CMOS-based 2T0C DRAM unit cells with the NN, NP, PN, and PP device types were designed with a unit cell area of 0.39 μm² for fair comparison, as shown in Figure 4. The unit cell layout was derived from the 4T2C eDRAM presented in [14]. It is noted that the unit cell area of the NN- and PP-type cells can be reduced because they share a common well (P-well for NN-type cell and N-well for PP-type cell), in contrast to the NP- and PN-type cells, which require separate N-well and P-well regions in each unit cell.

3. Comparative Analysis of Electrical Characteristics

Figure 5 presents the timing diagrams of the write, precharge, and read operations, and the corresponding V_SN curves in the 2T0C DRAM unit cells. After the write operation, a precharge voltage of 0.5 V (½V_DD) was applied to the write bitline (WBL), and a long read pulse was applied to RWL and RBL to evaluate the effect of repeated read operations and to measure the retention time. During the write operation, V_SN is coupled down in the NN- and NP-type cells and up in the PN- and PP-type cells to V_{SN_INIT}, mainly due to C_SN-WWL coupling. During the read operation, V_SN is coupled up in the NN- and PN-type cells and down in the NP- and PP-type cells by ΔV_{SN_READ} due to C_SN-RWL and C_SN-RBL coupling, respectively. The coupling effects during the write and read operations reduce the on/off current ratio, which was measured immediately after a read pulse.

Table 2. V_{SN_INIT}, ΔV_{SN_READ}, and on/off current ratio in 65 nm CMOS-based 2T0C DRAM unit cells.

Device Type		NN	NP	PN	PP
VSN_INIT	Data ‘1’	0.69 V	0.02 V	0.98 V	0.31 V
	Data ‘0’	0.02 V	0.65 V	0.33 V	0.99 V
ΔVSN_READ	Data ‘1’	0.12 V	−0.10 V	0.10 V	−0.11 V
	Data ‘0’	0.10 V	−0.10 V	0.10 V	0.09 V
On/off current ratio		28.04	3.84	4.36	46.70

summarizes V_{SN_INIT}, ΔV_{SN_READ}, and the on/off current ratio in the 65 nm CMOS-based 2T0C DRAM unit cells. Since the coupling direction of V_SN for data ‘0’ in the NP- and PN-type cells is the same during the write and read operations, V_SN for data ‘0’ in these cells approaches that for data ‘1’ more closely during a read pulse in contrast to the NN- and PP-type cells. As a result, the NP- and PN-type cells exhibit very low on/off current ratios, which can degrade inference accuracy in 2T0C DRAM-based CIM applications.

In the 65 nm CMOS-based 2T0C DRAM unit cells, the retention characteristics are mainly affected by the subthreshold leakage current in the write Tr. [9], which is determined by the magnitude of the voltage difference between SN and WBL (V_SN-WBL). After the write operation, a large V_SN-WBL for data ‘0’ in the NN- and PP-type cells and for data ‘1’ in the NP- and PN-type cells induces a significant subthreshold leakage current, leading to faster deterioration of V_SN. Additionally, unlike the NN- and PP-type cells, ΔV_{SN_READ} for data ‘1’ in the NP- and PN-type cells further increases V_SN-WBL during a read pulse, thereby accelerating the deterioration of V_SN. The retention times for data ‘1’ and data ‘0’ of the 2T0C DRAM unit cells were measured as shown in Figure 6. In this work, the retention time for data ‘1’ was defined as the time required for the initial I_READ for data ‘1’ (I_{READ_1}) to decrease by 2%. Conversely, the retention time for data ‘0’ was defined as the time required for the initial I_READ for data ‘0’ (I_{READ_0}) to increase by 2% of the initial I_{READ_1}. The retention times for data ‘1’ in the NP- and PN-type cells are significantly shorter than those in the NN- and PP-type cells. Considering both the on/off current ratio and the retention time, we conclude that the NP- and PN-type 2T0C DRAM unit cells are not favorable for CIM applications.

4. Investigation of V_SN Fluctuation

As shown in Figure 7, V_SN for data ‘0’ (red) in the NN- and PP-type cells fluctuates during the fluctuation interval immediately after a read pulse in contrast to that for data ‘1’ (blue). The reason the fluctuation interval in the PP-type cell is longer than that in the NN-type cell will be discussed at the end of this section. Since the V_SN fluctuation can considerably affect the retention time, we identified its cause. Figure 8a and Figure 9a illustrate the equivalent RC models of the NN- and PP-type 2T0C DRAM unit cells after the write operation for data ‘0’, respectively. As shown in Figure 8b and Figure 9b, V_SN varies by ΔV_{SN_READ} during the ramp-up and ramp-down of the read pulse. The magnitude of ΔV_{SN_READ} is determined by C_SN-RWL and C_SN-RBL coupling and further modulated by the coupling effect of the read gate oxide capacitance (C_OX). V_SN in the NN-type cell is coupled up due to C_SN-RWL coupling, whereas V_SN in the PP-type cell is coupled down due to C_SN-RBL coupling, which causes variations in the channel potential of the read Tr. (V_{READ_CH}) by +ΔV or −ΔV. Simultaneously, V_{READ_CH} in the NN-type cell increases, while V_{READ_CH} in the PP-type cell decreases according to the voltage division between RWL and RBL. As a result, V_SN is further coupled up or down through C_OX coupling. As shown in Figure 8c and Figure 9c, the ΔV component in V_{READ_CH} dissipates through I_READ and V_{READ_CH} saturates to the divided voltage between RWL and RBL. Subsequently, V_SN increases in the NN-type cell and decreases in the PP-type cell due to the subthreshold leakage current in the write Tr., as illustrated in Figure 8d and Figure 9d.

Consequently, the dissipation speed of the ΔV component through I_READ and the subthreshold leakage current can affect the V_SN fluctuation interval, and the fluctuation interval may become longer as both I_READ and the subthreshold leakage current decrease. To verify this possibility, we adjusted the body voltage (V_B) from 0 V to −1 V in the NN-type cell and from 1 V to 2 V in the PP-type cell, thereby reducing both I_READ and the subthreshold leakage current. Figure 10 shows the V_SN fluctuation intervals for data ‘0’ in the NN- and PP-type cells under different V_B conditions, where the intervals are extended by 13.8% and 533.4% by V_B adjustment, respectively. Since the hole mobility is lower than the electron mobility, both I_READ and the subthreshold leakage current in the PP-type unit cell are lower than those in the NN-type unit cell. Therefore, the fluctuation interval in the PP-type cell is longer than that in the NN-type cell under the initial V_B conditions, as shown in Figure 10a,b. Likewise, the fluctuation interval in the PP-type cell becomes significantly wider under the adjusted V_B conditions.

5. Conclusions

In this work, we investigated the electrical characteristics of 2T0C DRAM unit cells with the NN, NP, PN, and PP device types for CIM applications. In conclusion, the NN- and PP-type 2T0C DRAM unit cells are relatively suitable for CIM applications, whereas the NP- and PN-type 2T0C DRAM unit cells are not favorable due to their very low on/off current ratios and short retention times. We elucidated the cause of the V_SN fluctuation using the equivalent RC models and showed that the fluctuation interval can be extended as both I_READ and the subthreshold leakage current decrease under the adjusted V_B conditions. This work is anticipated to make a significant contribution to the analysis of silicon channel-based 2T DRAM.