1. Introduction
In the current era of mixed-signal system-on-chips, digital integrated circuit (IC) design has the benefit of relying on mature automated synthesis tools for automated design. In contrast, analog IC design is a time-consuming task that heavily relies on manual analysis by expert designers due to the lack of mature automated synthesis tools [
1]. This manual design process is a bottleneck in IC design, prompting the need for automated analog IC design technologies, which have garnered widespread attention.
The mainstream approaches to automation can be broadly categorized into two types: optimization-based and knowledge-based methods [
2]. The optimization-based approach utilizes optimization algorithms such as evolutionary algorithms [
3] to generate new device size results. However, due to the inherent complexity of both algorithms and circuits, this approach may result in significant time consumption or even physically infeasible solutions [
4].
The knowledge-based approach, on the other hand, relies on expert knowledge to develop automated design programs capable of generating valuable solutions. Currently, the mature approach is based on the
method [
5] with pre-computed lookup tables (LUTs) [
6], which achieves high design accuracy and has been successfully applied in the design of various circuits such as operational amplifiers, bandgap reference circuits, low-noise amplifiers, and others [
7,
8,
9,
10,
11].
The LUTs used in the
method are generated by simulating the device parameters using a simulation program with integrated circuit emphasis (SPICE) models with fixed step sizes in channel length (
L),
,
, and
. For parameter values that are not present in the LUTs, interpolation can be utilized to predict the values. However, high-precision LUTs require more finely scanned parameter step sizes, leading to increased storage space usage [
12]. Taking into account the number of model parameters and device types, the large storage space requirements and the need to load them into memory for access entail significant hardware resource costs, rendering this design method impractical.
With the continuous advancements in machine learning (ML), researchers have started exploring the possibility of constructing device parameter models using ML models. Habal et al. [
13] developed a simple quadratic polynomial as a feature engineering input and proposed a neural network with a single hidden layer for predicting
. In the work of Habal et al., the prediction errors of the ML model were approximately controlled within 3%. However, predicting a single parameter of a transistor in a specific technology appears to be insufficient.
Yang et al. [
14] conducted a thorough analysis of the impact of different activation functions in a multi-layer perceptron model for predicting the parameters of various transistors. They proposed a neural network architecture with three hidden layers and found that the inverse square root unit (ISRU) activation function provided the best training results. The mean absolute percentage error (MAPE) for the negative channel-metal-oxide-semiconductor (NMOS) was reduced to 1.38%, but the errors for the positive channel metal-oxide-semiconductor (PMOS) ranged between 6.47% and 10.69%. Although these studies demonstrate the potential of multi-layer perceptron models instead of LUTs, the challenge lies in determining the optimal hyperparameters of the neural network.
Ho et al. [
15] proposed an algorithm that utilizes genetic algorithms and multi-layer perceptrons to predict current (
) using an evolved neural network. Their work demonstrated that the prediction accuracy of the evolved neural network was superior to that of a conventional multi-layer perceptron. This research provides some inspiration for our current work as it highlights the potential of using genetic algorithms to optimize neural networks and improve prediction accuracy.
Wang et al. [
16] introduced a direct-current (DC) simulation-based neural network model called the DC model. This model utilizes graphics card-accelerated neural networks to capture the non-linear relationship between specific DC parameters and performance metrics. By substituting certain simulation tasks, it significantly reduces time consumption while maintaining a high level of accuracy. Qi et al. [
17] proposed a knowledge-based neural network approach that segregates geometric variables from other input variables. The geometric variables are modeled using physics-based analytical equations, while the remaining variables are represented by an artificial neural network. This methodology was verified using the BSIM6 (BSIM, a simulation model developed by the UC Berkeley) model and demonstrated good agreement between the model predictions and experimental results. Similarly, Wang et al. [
18] conducted similar work by modifying the BSIM model for low-temperature conditions in a 180 nm technology. They introduced an optimization model based on the backpropagation neural network prediction to compensate for the low-temperature effect. A common characteristic of the aforementioned approaches is the utilization of neural networks to substitute certain aspects of physical modeling, thereby accelerating scientific computations.
Fu et al. [
19] presented a model based on the backpropagation neural network for the prediction of NMOS performance parameters. The modeling process involved employing substrate bias, substrate impurity concentration, oxide thickness, and adjusted implant doping concentration for threshold voltage as independent variables while considering threshold voltage and others as dependent variables. Following the training of the model, none of the final prediction variables exhibited an average percentage error exceeding 1.5%.
Wei et al. [
20] developed a precise artificial neural network model that encompasses the complete range of drain currents. By employing the Latin hypercube sampling algorithm, they achieved a substantial reduction in the training data requirement without significantly compromising the quality of the fitting outcomes. This approach effectively mitigated training overhead. The experimental findings demonstrate that the proposed artificial neural network model exhibits excellent fitting capabilities in the 180 nm technology.
Most of the above works focus on the parametric modeling of devices, but less work applies the proposed models to actual circuit design. In addition, the technology nodes involved in device parameter modeling focus on one or two technologies, and the effectiveness of the model on more technology is unknown.
This paper introduces a new deep neural network (DNN) architecture that has multiple DC outputs for modeling complementary metal-oxide-semiconductor (CMOS) device parameters. This DNN model requires only four inputs (, , , and L) and can produce 14 outputs simultaneously, including , , , , , , , , , , , , , and . The suggested DNN architecture comprehensively models the device parameters to prevalent process design kits (PDKs) TSMC 40 nm (T40), TSMC 65 nm (T65), TSMC 180 nm (T180), and SMIC 180 nm (S180), respectively. Compared with the traditional LUTs, the proposed DNN models for each PDK occupy less storage space and have high-accuracy prediction performance. Moreover, this paper leverages the DNN models within each PDK, in conjunction with parameters, to achieve circuit migration design. By employing circuit multiplexing, the efficiency of analog integrated circuit design is enhanced.
The structure of the paper is organized as follows:
Section 2 introduces the DNN architecture and outlines the necessary metrics requirements. In
Section 3, the dataset acquisition process is described.
Section 4 discusses the automated design method based on the
parameter and DNN models.
Section 5 presents the experimental results. Finally,
Section 6 is the conclusions.
3. Data Sampling
The DNN models are trained using four PDKs: T40, T65, T180, and S180, respectively. The respective ranges and step sizes of the variables used to generate each PDK dataset are presented in
Table 3. For T40 and T65, the range of
and
is set from 0.1 V to 1.2 V, with a step size of 0.02 V for both variables. Similarly, the range of
is from 0 V to 1.2 V, with a step size of 0.02 V. On the other hand, for T180 and S180, the range of
and
is extended from 0.1 V to 1.8 V, and the range of
spans from 0 V to 1.8 V, with a consistent step size of 0.02 V for both variables. The channel length (
L) values vary from 0.5 µm to 6 µm, with a step size of 0.2 µm, while the channel width (
W) remains fixed at 5 µm across all PDKs.
Table 4 provides detailed information regarding the generated datasets, including their sizes and the parameters they encompass. Each dataset comprises 14 device parameters, namely
,
,
,
,
,
,
,
,
,
,
,
, and
. The dataset sizes for T40 and T65, which were saved in pickle format, were 1.4 GB space each (including both PMOS and NMOS devices). Similarly, the dataset sizes for T180 and S180 were 4.4 GB space each (including both PMOS and NMOS devices).
Prior to training, the logarithm of the 14 target parameters to be predicted was taken, and the dataset was normalized to eliminate any significant variations among the target parameter data. Subsequently, the dataset was divided into a training set, a validation set, and a test set, following a ratio of 0.81:0.14:0.05, respectively.
4. Sizing Method
Figure 2 illustrates the design flow used in this work. Specifically, the flow begins by specifying the design variables for the circuit, which include the
and
L for each transistor, as well as the passive device resistor-capacitor (
) values. Subsequently, the design points for
,
,
for each transistor are determined using the mapping of
and
, along with the application of Kirchhoff’s voltage law (KVL) to the circuit. These design points (
,
,
and
L) are then input into
to obtain 14 DC parameters.
The obtained 14 DC parameters serve two purposes: firstly, they are combined with the circuit performance equations to predict the performance of the circuit. Secondly, they are utilized to solve for the channel width (W) of each transistor based on the method. To expedite this process, parallel computing techniques are employed. Upon traversing all the design points, the values of L, W, R, and C alongside the corresponding circuit performance are stored in the . Finally, the sizing results are outputted according to the performance specifications requirements.
To ensure the efficacy of the design, this study incorporates a parameter known as voltage saturation margin (
). This parameter determines whether a transistor operates in the saturation region by evaluating the condition
. Any design point that includes transistors failing to satisfy the aforementioned condition is discarded. The working details of the
parameter are shown in
Figure 3.
The reuse of analog circuit designs represents a pivotal strategy for enhancing the efficiency of IC design. The approach employed in this paper for implementing circuit migration design is founded upon the design process outlined in
Figure 2. Specifically, it entails leveraging distinct DNN models associated with different PDKs to generate the circuit design database corresponding to each respective PDK. Subsequently, the final device sizes are determined based on the performance requirements.
It is noteworthy that the methodology proposed in this paper involves traversing the design space and recurrent utilization of the DNN models for parameter prediction, which constitutes a significant portion of the overall process time. Consequently, the time-consuming nature of the model prediction phase holds great importance as it directly impacts the efficiency of circuit design database generation.
6. Conclusions
This paper presents a novel approach for training DNN models to replace the LUTs in the method, thereby significantly reducing storage space requirements. The proposed method focuses on training DNN models for NMOS and PMOS device parameters in prevalent IC design technologies, namely T40, T65, T180, and S180. By employing DNN models, the storage space requirements can be reduced by at least 99% compared to LUTs. Furthermore, the DNN models demonstrate high prediction accuracy, with an average percentage error of less than 1%. In terms of prediction time, the DNN models outperform LUTs by reducing the time overhead by 91.46% when predicting an equivalent number of parameters.
Additionally, this paper introduces an automated porting design approach for analog circuits that combines the DNN models with the design method. The objective is to facilitate circuit design reuse across different technologies. The proposed method is validated through migration design experiments involving folded cascode amplifiers and Miller two-stage amplifiers in the mainstream technologies of T40, T65, T180, and S180. The experimental results demonstrate that utilizing the DNN models can reduce the time overhead by 40% compared to using LUTs. Furthermore, the pre- and post-simulation results of the circuit confirm that the proposed method enables the automated design of circuits with identical specifications across different technologies.
Based on the experimental results, it was observed that to achieve the same performance specifications, the sizing of operational amplifiers in T40 technology is relatively larger than that of T65 technology for both circuit architectures. Therefore, it is recommended that for advanced technology circuit design, the performance targets can be relaxed to obtain smaller circuit sizes. Alternatively, one may consider replacing the architecture to meet the desired specifications.