Global TCAD Solutions

Predictions based on Physics

Expanding TCAD Simulations from Grid to Cloud

In one of our 2015 SISPAD papers, we analyzed distribution, execution and performance of typical TCAD simulations on traditional grid-based systems versus cloud systems. As turned out, a hybrid approach is the most efficient and cost-effective solution in many cases.

The necessary algorithms and interfaces are included in GTS Framework; the integrated Design Of Experiment (DOE) and Optimizer modules split simulation jobs for optimal performance in a distributed environment — the latter being especially useful for parameter fitting / inverse modeling.

Physical Device Simulation in a Multi-Host Environment

Power and Allocation of Resources

Grid-based systems, such as Sun Grid Engine (SGE) or Load Sharing Facility (LSF) are pervasive in the TCAD domain, and the used nodes are usually highly optimized for the required kind of computations. In contrast, typical cloud solutions are less optimized, but offer the ability to create additional computing hosts on-demand, tapping into a virtually unlimited pool of resources. This allows to allocate huge amounts of computational power just when needed, adding a major boost for time-critical projects — which can translate to a significant advantage in business.

Cost Structure

The cost structure of used cloud resources is typically different from the one of an on-premise grid-based system. On popular cloud services, billing granularity is one hour of node uptime, which can amount to largely varying total cost depending on the runtime of the submitted tasks and on the used scheduling algorithm (for details, refer to our paper below).

Table 1: Cost outline for grid / cloud / hybrid system

Most Efficient: A Hybrid Approach

Since clould resources are typically more expensive than using on-premise grid systems, we suggest to evaluate a hybrid apporach. GTS JobServer's SmartCluster technology allows transparently using grid and cloud systems at the same time for each submitted simulation job – i.e. it pefers local grid ressources, and scales out to the cloud service during load peaks. In addition, GTS JobServer has a special heuristic algorithm which considers both the billing intervals of the cloud service as well as the expected / actual runtime of the submitted tasks to provide optimal overall job completion times at minimal cost. Table 1 shows a cost outline for the below example 2.

Fig.1: UI showing the running simulation to determine Vth,lin and Vth,sat

Ex. 1: Statistical BTI Degradation of FinFETs

Due to the small device size, the doping atoms cannot be assumed to form a continuum [Sano2001], and the position of each individual doping atom has a decisive influence on the transistor’s properties. Therefore, an array of virtual devices is generated, each with a different random distribution of discrete dopants.

We examine the current/voltage characteristics for the linear and the saturated operation regime of the transistor. The threshold voltages for both regimes, Vth,lin and Vth,sat, are determined for each specific doping configuration. GTS JobServer allows for easy task setup and shows job progress and postprocessing results in a single view. An exemplary view of such a simulation is shown in Fig.1.

Fig.2: GTS JobServer, sub-tasks
Fig.3: BTI-induced Vth shift

Ex. 2: Distributed Nano-Device Simulation

The subband-Boltzmann transport equation (BTE) is solved on slices along the channel region to determine effective mobility [Karner2015]. GTS JobServer takes care of multi-host parallelization of the subband and scattering rate calculation: Per request by the device simulator, mobility is calculated on device cuts in sub-tasks, see Fig.2.

The shift of the threshold voltage Vth is an important key figure for the transistor’s performance in an integrated circuit. A key figure is the BTI-induced Vth shift, where a small change in oxide trap distribution can cause considerable changes in Vth under device stress conditions [Cheng2010].

Large samples of trap/dopant configurations need to be investigated to achieve reasonable statistics on device reliability. A typical result of such a reliability/variability simulation (sample size 480) can be seen in Fig.3.: At the top, the BTI-induced Vth shift in an ensemble of n-FinFETs with randomly positioned oxide traps is shown; the lower plot shows ∆Vth distribution after a stress time of 105 s.

Fig.4: DOE in GTS Framework
Fig.5: Resulting CV curves

Ex. 3: Design of Experiment – Process Splits

Here, we study the effect of changes in a structure parameter (like the threshold doping) on the CV curve of a 65nm NMOS FET. We created the device from a template supplied with GTS Framework, and set up the simulation for the CV curve (f=1GHz, step VG from -2V to +2V, EW=-0,56V). Next, we used set up the variation of the threshold doping, which is done easily in the DOE (Design of Experiment) Script tool included in GTS Framework. Fig. 4 shows the dialogs for specifying the parameter that should vary. Of course, there can be more parameters to vary, and one can choose among various pre-defined DOE schemes (like full-factorial, full-grid, and various central composite designs) or design a custom experiment.

Obviously, parameter variation can largely benefit from parallel task execution, since the sub-tasks are independent from each other.

Fig.5 shows the resulting CV curves, for threshold dopings ranging from 0 to 1.2E19/cm3 with a step size of 4E18/cm3.

This is also a very basic example of using the DOE Scripting tool in GTS Framework – for more details, please see the DOE, Optimizer, Scripting tutorial, which explains each step in detail.

Ex. 4: Optimizer – Parameter Inverse Modeling

Fig.6: Optimized CV curve

Usually measurements are used to calibrate a TCAD model. In this simple example, we used GTS Framework's optimizer to fit a given CV curve, by modifying two structure parameters: threshold dopant and oxide thickness. For optimizations, GTS JobServer automatically schedules the sub-tasks for optimal parallelization, which provides a significant performance boost over serial execution. Since task scheduling is fully transparent in GTS JobServer, also such tasks can be scaled out to a cloud at load peaks, if desired.

Fig. 6 shows the history of optimization steps of the CV curve (green) approaching the measured curve (orange).

This example is explained in more detail the DOE, Optimizer, Scripting tutorial (section 1.3).


[Sano2001]: N. Sano and M. Tomizawa, “Random dopant model for three-dimensional drift-diffusion simulations in metal-oxide-semiconductor field-effect-transistors”, Applied Physics Letters, vol. 79, no. 14, pp. 2267–2269, 2001. [Online]. Available:

[Karner2015]: M. Karner, Z. Stanojević, F. Mitterbauer, C. Kernstock, and H. Demel, “Bringing Physics to Device Design – a Fast and Predictive Device Simulation Framework” in 2015 Silicon Nanoelectronics Workshop (SNW), 2015, pp. 75–76. Also see News: GTS @ VLSI, IRPS, SNW.

[Cheng2010]: Hui-Wen Cheng, Fu-Hai Li, Ming-Hung Han, Chun-Yen Yiu, Chia-Hui Yu, Kuo-Fu Lee, and Yiming Li, “3D device simulation of work function and interface trap fluctuations on high-K / metal gate devices” in Electron Devices Meeting (IEDM), 2010 IEEE International, Dec. 2010, pp. 15.6.1–15.6.4.