 Research
 Open access
 Published:
A holistic fast and parallel approach for accurate transient simulations of analog circuits
Journal of Mathematics in Industry volumeÂ 7, ArticleÂ number:Â 12 (2017)
Abstract
The accurate analog simulation of critical circuit parts is a key task in the R&D process of integrated circuits. With the increasing complexity of integrated circuits it is becoming cumulatively challenging to simulate in the analog domain and within reasonable simulation time. Previous speedup approaches of the SPICE (Simulation Program with Integrated Circuit Emphasis) analog circuit simulator included either solver improvements and speedup or model order reduction of the semiconductor devices.
In this paper we present a comprehensive approach to significantly speedup a SPICEbased analog circuit simulator while keeping the singlerate characteristic of time domain simulations. The novelty of our approach consists in the combination and extension of existing approaches in a unique way, enabling fast transient SPICElevel simulations. The main component of our approach is the circuit partitioner that combines relevant aspects from circuit theory and linear algebra in a unifying way. This enables the construction of an efficient and parallel BBD (bordered block diagonal) solver. Furthermore, this BBD structure allows for intrinsic model order reduction of the partitions during the Newton iteration, transforming the Newton method to a QuasiNewton method.
For midsized and largesized circuits our BBD approach leads to significant sequential and parallel accelerations of transient simulations. Additional speedup can be gained from our blockbypass strategies exploiting the latency in the partitioned circuit. Altogether our approach leads to a speedup of up to two orders of magnitude compared to the stateoftheart KLU solver while maintaining SPICElevel accuracy.
1 Introduction
The introduction of the first SPICEÂ [1] simulator revolutionized the design of electrical circuits. The ability to simulate the circuit in various conditions and scenarios, before the circuit is actually built, is crucial for fast and cheap circuit development.
Stateoftheart commercialÂ [2, 3] and opensourceÂ [4, 5] SPICE simulators offer numerous analyses in different physical domains and for various circuit types, helping the engineers to analyze the behavior of the circuit under different conditions. One of the most basic and workhorselike analyses of SPICE simulators within a commercial R&D environment is the transient analysis which computes the behavior of a circuit in the time domain up to a specified time point. Although one might consider the transient analysis as consecutive operating point analyses in the time domain, we consider the transient analysis as a separate analysis. Anyway the presented methods exploit some of the particular aspects of the transient analysis which could not be applied within an operating point analysis.
The computational overhead of a transient analysis is increasing with the circuit size and with the number of time steps required during the discrete time integration. With constantly growing circuit complexity the transient simulation poses a significant bottleneck in the circuit design and verification pipeline. Hence, we focus in this paper on a holistic approach to improve the performance of the transient analysis in a general SPICE simulator, while maintaining the same level of accuracy. The proposed methods from this paper were implemented within the frame of the analog inhouse circuit simulator of Infineon called TITANÂ [6â€“8]. However, our approaches are of general character and can be exported not only to other analog simulators but also to more general nonlinear transient problems.
SPICE simulators like TITAN use the modified nodal analysis (MNA) to build up the mathematical representation of the circuit. This formulation is mainly based on Kirchhoffâ€™s current law stating that the balance of all incoming and outgoing currents in a node must sum up to zero. The additive contribution of each individual device is specified by its characteristic equation which might have static and/or dynamic contents. The unknowns of the mathematical system essentially represent the node voltages; however, for each voltagedefining element such as voltage sources, an additional variable has to be introduced which represents the contribution of its branch current in the Kirchhoffâ€™s current law formulation. Altogether this approach leads to a system of differentialalgebraic equations of the following general form
Here \(q, f \in\mathbb{R}^{n}\) are the charge vector and the resistive current vectors, whereas \(s \in\mathbb{R}^{n}\) is the stimulus vector. The vector \(x(t)=(v(t),i(t)) \in\mathbb {R}^{n}\) represents the unknowns of the system: the node voltagesÂ v and the MNA branch currentsÂ i, respectively. As a starting point for the transient analysis the values of \(x(0)=(v(0),i(0))\) are given or computed by an operating point analysis. Starting at \(t=0\) the goal of the transient analysis is to compute the unknown values for \(t \in[0,T]\), where T represents the stop time of the simulation. For more details on the circuit modeling we refer toÂ [9].
The DAE systemÂ (1) is discretized for time points \(t_{k} \in(0,T]\), \(k=1, \ldots, M\). Considering the backward Euler formula as a simple representative of an implicit integration scheme, the discrete form of (1) at time point \(t_{k}\) is
with the time step \(h_{k}=t_{k}t_{k1}\).
For each time point \(t_{k}\) a system of nonlinear equations, derived from the discretized DAE systemÂ (2), needs to be solved with the unknown vector \(x_{k}=x(t_{k})\):
More generally, an implicit integration method takes the form
with the integration coefficient \(\alpha_{k} \in\mathbb{R}\) and the history information \(\beta_{k} \in\mathbb {R}^{n}\) which are both constant for the current time point. Applied to the differentialalgebraic systemÂ (1), the nonlinear system at time point \(t_{k}\) for the unknown \(x_{k}\) reads
Applying the NewtonRaphson method transforms this nonlinear system into a sequence of linear problems: The linearization of (4) at the current value of \(x_{k}\) specifies the Newton correction \(\Delta x_{k}\) as the solution of the linear system
We denote by \(C(x_{k})\) and \(G(x_{k})\) the Jacobian matrices of the charge and of the resistive current vectors. Then \(A(x_{k}) = \alpha_{k}\cdot C(x_{k}) + G(x_{k})\) represents the combined system matrix. The residual on the righthand side of (5) is referred to as \(r(x_{k})\), so that the Newton correction \(\Delta x_{k}\) is the solution of the linear system \(A(x_{k}) \Delta x_{k} =  r(x_{k})\).
The linear systemÂ (5) is the foundation of the Newton method which is required for each time step k. The Newton loop is the most inner loop (see FigureÂ 1) within the transient analysis and represents the largest portion of the computational task within a transient simulation.
In order to speed up the transient simulation while maintaining the SPICElevel accuracy, the algorithm on FigureÂ 1 should be significantly improved, while keeping the singlerate and implicitly coupled characteristic of the algorithm.
Abandoning the singlerate principle or the implicit characteristic of the algorithm on FigureÂ 1 results in a fastSPICE algorithmÂ [10, 11] that especially for large circuit have the potential to be 10100 times faster than their SPICE counterparts but are considerably inaccurate and with default settings, without circuit specific settings, there is a high probability of producing wrong results. However, for large circuits and long transient simulations, the only feasible analog simulation is through such fastSPICE algorithms.
Partitioning the circuit is a fundamental element in most of the fastSPICE acceleration techniques. In one of the fastSPICE approachesÂ [12, 13] the partitions are solved implicitly but are coupled explicitly in a multirate manner. In this case the static partitionerÂ [12] of the circuit must capture each feedback loop within the same partitionÂ [13] and decouple the circuit along weak capacitive connections. A similar approach is presented inÂ [14], where the feedback loops are also included in the same partition, but the partitions are overlapping and coupled explicitly by multiplicative or additive Schwarz iterations resulting in an accurate and fast simulation especially for RC dominated circuits. Another fastSPICE approachÂ [15] uses the hierarchical circuit description in the input netlist^{Footnote 1} to partition the circuit by capturing the repetitive elements in the circuits. If many instances of a given element share the same state, then significant computation can be saved during step (1.4) in FigureÂ 1. Furthermore, depending on the coupling strengthÂ [15], selected partitions can be coupled explicitly. A similar approach is introduced inÂ [16] with emphasis on efficient and parallel computing. The method presented inÂ [17] also uses the hierarchical netlist structure for partitioningÂ [6] and applies the multirate time integration for the resulted partitioning, resulting in a considerably shorter simulation time but also in unpredictable accuracy. For a comprehensive overview of fastSPICE techniques we refer further toÂ [10, 11].
Another important approach to further speedup the simulation is to apply modelorder reduction to the circuit used as input for the simulator such that accuracy is not compromised and simulation time is drastically reducedÂ [10, 18, 19]. These methods are also named as SPICEin, SPICEout network reduction methods, and they work especially well with circuits dominated by passive elementsÂ [18, 19]. In an industrial environment they are typically applied before the actual SPICElevel simulations are started.
In order to deliver reliable SPICE accuracy, preserving the algorithmic framework in FigureÂ 1 is an important aspect of the SPICE acceleration techniques. The main focus of these methods is to speed up and parallelize the inner loop of the algorithm, steps (1.4)(1.7) in FigureÂ 1. The direct linear solvers in SPICE simulators, depending on the nature of the circuit, scale \(O(n^{1.3\text{}1.8})\) with the sizeÂ n of the circuitÂ [11]. Constructing the matrix and the righthand side in step (1.4) is done with linear complexity \(O(n^{1})\), but the constant factor of the complexity function is of order 10^{3}10^{5}, since the industrial semiconductor models typically have numerous nonlinear equations to evaluate. All other steps ((1.6) and (1.7) in the inner loop) have linear \(O(n^{1})\) complexity and only require a couple of operations per MNA variable. Therefore for large circuits the solver step (1.5) is the most dominant part of the Newton loop and was the subject of several research work. In several publicationsÂ [20â€“23] the authors present various methods to speed up and to parallelize the linear solver step in the inner loop. These approaches are limited not just by Amdahlâ€™s law for parallelization, but they also have limited speedup capability due to their pure linear algebra view on the problemÂ [10]. Even the approaches that partition the circuit on the pure linear algebra viewÂ [24] of the problemÂ [10, 11] have limited speed up capacity for different circuit types and sizes.
For midsized circuits,^{Footnote 2} the most computeintensive part of the inner loop is the evaluation of the semiconductor models during the setup of the linear systemÂ (5), step (1.4) in FigureÂ 1. Hence, other approaches target the evaluation of the semiconductor devices by either accelerated and parallelized evaluationÂ [5, 25, 26] or by model order reduction such as table modelsÂ [27] of the complex semiconductor models.
Focusing only on one single step of this inner loop results in significant speedup either for RC^{Footnote 3} or semiconductordominated^{Footnote 4} circuits of either large or midsized circuits. For small^{Footnote 5} and midsized^{Footnote 6} circuits the semiconductor model evaluation and building the Jacobi matrix is usually the most dominant part. For largesized circuits,^{Footnote 7} most of the computational workload is in the linear solver. In addition to the size n of the circuit, it is also relevant which type of devices are dominating the circuit: If the circuit is dominated by parasitic resistors and capacitances and their number is usually considerably larger than n, then the Jacobi matrix becomes more dense and the solver more computationally dominant.
1.1 Structure of this paper
In this paper we present a comprehensive approach to speed up and parallelize all the computational intensive steps of the inner loop in FigureÂ 1, while maintaining the singlerate characteristic and thus the reliable SPICE accuracy of the algorithm.
The starting point of our approach is a circuit partitioner. Our approach of partitioning uses, extends, and combines existing partitioning approaches. Similar to other domain decomposition approachesÂ [6, 28] our partitioner minimizes the number of connection nodes between the partitions, but in addition it also makes sure that the fillin rate of the resulting coupling system is limitedÂ [24] and that all partitions can be evaluated and solved in a fully parallel wayÂ [6]. This partitioner is presented in the first section of this paper.
In the second section of the paper, we present the resulting BBD matrix data structure and the hybrid solver that extends the approach presented inÂ [24].
The third section of this paper describes the partition bypass approach that extends the BBD solver and the partition evaluation process. By skipping parts of step (1.4) and step (1.5) for converged partitions of the inner Newton loop in FigureÂ 1 significant computation can be skipped, yet maintaining the same numerical precision of the simulation.
The final section of this paper presents the numerical results and simulation time comparison of our approach. We measure the run time and check the accuracy of our implementation for a large range of circuit types and circuit sizes. Thereby we demonstrate that our approach results in significant parallel and sequential speedup compared to classical SPICE algorithms and also delivers reliable accuracy for all circuits that we simulated.
2 Methods
2.1 Circuit partitioner
In most comprehensive SPICE and fastSPICE acceleration approaches, the static or dynamic partitioner plays a central role. In our approach, since we keep the singlerate principle of the simulation and also for sake of simplicity, we only considered a static partitioner that divides the circuit in an early setup phase into a predefined number of partitions. In contrast to a static partitioner, dynamic partitioners repartition the circuit during the transient analysis based on the current state and activity of each circuit partÂ [15].
Due to the static nature of our partitioner it is crucial that all relevant performance aspects and constraints are considered already in the setup phase. We group the devices from the circuit into separate partitions as it is illustrated in FigureÂ 2 for a simple circuit with only two partitions. The resulting partitions in FigureÂ 2 have no electrical meaning because the partitioner minimizes the coupling nodes regardless of the subcircuit structure of the input circuit. Similar to common circuit partitioning approachesÂ [6, 11, 17, 29], the devices are grouped such that the number of coupling nodes are minimized. This objective can be achieved with a general graph partitioning algorithmÂ [30â€“32], where the devices are the nodes in the graph and the matrix entries in A (see (5))^{Footnote 8} of the circuit represent the edges in the graph. Another approach is to transform the circuit into a hypergraph, and use an appropriate partitioning algorithmÂ [29].
After grouping the devices into partitions and ordering the rows and columns of A according to these partitions, the system matrix will have the BBD structureÂ (6).
The block index c represents the coupling part of the p partitions in the circuit. One main objective of the partitioner is to create a matrix structure that can be built and partially solved independently by each partition. In this way, the coarsergrained parallelism can be realized significantly improving the parallel performance also for mid and smallsized circuits. Since this BBD structureÂ (6) is resulting from the device grouping partition approach, this also means that \(A_{c,c}\) is the only matrix block that needs synchronization for writing in a multithreaded mode. Fortunately, this write conflict can be simply solved by splitting the coupling system \(A_{c,c}\) into the contributions from each partition, i.e.,
After splitting the coupling matrix \(A_{c,c}\) the global system A can be written also in this split formÂ (7) with the following matrix structure:
Analog to FigureÂ 2 we denote the partition matrices by \(A = A_{P1} + A_{P2} + \cdots+ A_{Pp}\), where the global system matrix is the sum of all partition matrices.
Applying the same splitting to the righthandside vector, we get p separate righthandside vectors \(r(x_{i})= r_{P1} + r_{P2} + \cdots+ r_{Pp}\):
Given the splittings (8) and (9), and the condition that each device must be assigned uniquely to one partition, there is no writing conflict for devices from different partitions, hence the partitions can build their matrices (step (1.4) in FigureÂ 1) in a fully parallel and unsynchronized manner. In the upcoming section about partition bypass, we present further aspects of this matrix and righthandside separation, which will turn out to be beneficial also for sequential simulations.
The next crucial objective of our partitioning approach is to limit the fillin entries in \(A_{c,c}\) which will result from the LU solving process. A given matrix \(A_{Pi}\) of partition i is partially LU decomposed except the \(A_{c,c}^{(i)}\) part. The solving step of the BBD matrix is presented in detail in the next section.
In previous approachesÂ [6, 11] the authors point out that if these fillins are not controlled then for only 10^{3}10^{4} coupling nodes solving the coupling system \(A_{c,c}\) of the BBD matrix becomes a bottleneck or even unfeasible. InÂ [24], the author uses a fillin minimization technique by analyzing the elimination tree of the global BBD matrix and identifying the coupling nodes such that the fillin entries are minimized. However, the resulting partitioning cannot be built independently by the devices, since this approach considers only the matrix view of the circuit.
Since it is difficult to compute the fillin rate at the circuit device level, we analyze the fillin rate of the coupling system in a second step after the device grouping, once the fillin entries in \(A_{c,c}^{(i)}\), \(i=1,\ldots,p\) can be computed. In this second step, for each row of \(A_{i,i}\) and \(A_{i,c}\) we compute how many fillin entries would be inserted into the matrix blocks \(A_{c,i}\) and \(A_{c,c}^{(i)}\). If the number of fillins for a row exceeds a threshold in the range of \([ 10^{3}, 10^{4} ]\), then the row is moved from \(A_{i,i}\) and \(A_{i,c}\) to the coupling part \(A_{c,i}\) and \(A_{c,c}^{(i)}\). In our empirical tests it turned out that a constant threshold of 1500 increases the number of coupling nodes only marginally but decreases the number of fillins drastically in the accumulated \(A_{c,c}\). The value of this threshold can be explained by the structure of the system matrix, where most of the matrix rows have less then 10 entries, therefore they cause marginal fillin entries. According toÂ [24] and to our experience, there are only proportionally few rows that cause a significant number of fillins, and these rows are detected by this threshold value and are moved to \(A_{c,c}\).
The final objective of our partitioning approach is to ensure that all block matrices can be LU decomposed by a static pivoting solver. The fillin minimization reordering does always symmetric row and column swapping, so that the diagonal elements remain on the diagonal. Node voltage MNA variables can be used for static pivoting, since they have nonzero diagonal entries. However, the branch current MNA variables which are required by voltage sources and inductors have no diagonal entries in all analyses. Therefore they require a neighboring node voltage MNA variable for pivoting. This row swapping between these two MNA variables must be possible within one partition or within the coupling part of the system. In this last step of the partitioning, it is ensured that all current MNA variables and their neighbor node voltage MNA variables are either in the same partition block matrix \(A_{i,i}\) or in the coupling system \(A_{c,c}\). For this reason, if necessary, further MNA variables are moved to the coupling system. In other words, we ensure that all current paths are contained completely either in a unique partition matrix or in the coupling matrix. In this way it is always ensured that the diagonal entries will stay nonzero during the blockwise Gaussian eliminations.
The novelty of our circuit partitioning approach is that we consider various aspects of the problem, that all enable at the end a fast and parallel simulation of midsized and large circuits. At the end of this section as a summary we list all the objectives that represent the core of our partitioning approach.

1.
Parallel building and solving of each partitionâ€™s matrix.

2.
Minimize the number of coupling nodes.

3.
Minimize the fillin rate in the coupling system.

4.
Ensure for all partitions the solvability with static pivoting LU solvers.
These objectives are achieved by the following three consecutive steps that form our novel partitionerâ€™s pseudo algorithm:

1.
Group the devices into partitions with the graph partitioning algorithmÂ [30] such that the coupling nodes are minimized.

2.
Apply the fillin threshold of 1500 to minimize the fillin rate in the coupling system.

3.
If necessary, ensure for all partitions the solvability with static pivoting LU solvers by moving MNA variables into the coupling part.
2.2 Solving the BBD system
In this section, we present in more detail our approach to solve the BBD system. For more details on the mathematical background of the BBD systems and Schur complement we refer toÂ [11, 24, 33]. By permuting the rows and columns according to partition nodes \(x_{i}\), \(i=1,\ldots,p\) or coupling nodes \(x_{c}\), the BBD system matrix A and righthand side r take shape as
The solution of each partition from (10), assuming \(x_{c}\) is known, is given by \(x_{i} = A_{i,i}^{1}(r_{i}  A_{i,c} x_{c})\), whereas the solution for the coupled system is determined by
Using the splitting of the partitioner for parallel matrix and righthand side building, introduced in (8) and (9), the coupled systemÂ (11) is transformed to
Each summand in the sums of (12) is the contribution from one partition, and only in the final step they need to be summed sequentially. Using the abbreviations
the coupled systemÂ (12) can be written as
yielding the coupling system \(S x_{c} = s\) which needs to be solved before all \(x_{i}\), \(i=1,\ldots,p\) can be determined.
The contributions \(S_{i}\) and \(s_{i}\) of partitions \(i=1,\ldots,p\) to the coupling matrix S and righthand side s are computed by a partial inplace block LUdecomposition of each partitionâ€™s matrix, i.e.,
The matrix \(S_{i}=A_{c,c}^{(i)}  A_{c,i}U_{i,i}^{1}L_{i,i}^{1}A_{i,c}\) will be the direct result of the partial block LUdecomposition, and the righthandside contribution is given by the subsequent partial forward substitution of the resulting LUdecomposition with the given righthand side \(s_{i} = r_{c}^{(i)}  A_{c,i}U_{i,i}^{1}L_{i,i}^{1}r_{i}\) for \(i=1,\ldots,p\). Due to the criterion 4 of the partitioner it is ensured that the partial LUdecomposition is made with inplace and static pivoting solver, resulting in maximal numerical performance. Referring to FigureÂ 1, we also point out that with our BBD approach not just step (1.4) but also partially step (1.5) can be computed in parallel. The first synchronization point in the Newton loop is the building and solving of the coupling system \(S x_{c} = s\).
Subsequent to the coupling system solution and in accordance to the partial LU decompositionÂ (13), a partial backward substitution is necessary to compute the unknowns \(x_{i} = U_{i,i}^{1}L_{i,i}^{1} (r_{i}  A_{i,c} x_{c})\), \(i=1,\ldots,p\). Once \(x_{c}\) is known, this step can be computed in a parallel and unsynchronized manner.
FigureÂ 3 summarizes the presented steps to build and to solve the BBD system. The three steps on FigureÂ 3 represent steps (1.4) and (1.5) from FigureÂ 1 which is the most computational intensive part of the Newton loop within transient simulations. The sequential part of FigureÂ 3 is grouped into step (3.2) that needs to be addressed for both parallel and sequential computational performance.
2.3 Solving the coupling system
Building the matrix S and corresponding righthand side s is a sequential but linearly complex task. Since the matrix S is more dense than the partition matrices \(A_{P_{i}}\), \(i=1,\ldots,p\) [11, 24], solving this coupling system becomes more computational complex than building the matrix and righthand side. Therefore, in the following subsection, we focus on our approach to solve the coupling system that is a crucial element in step (3.2) of FigureÂ 3.
The criterion 3 of the partitioner ensures that the number of fillins in the matrix are reduced, but they are overall significantly higher than for the individual partition matrices.
For midsized circuits, if the number of coupling nodes is less than a few hundreds and a direct solver does not produce significant additional fillins, a direct solver works the best also for the coupling solver. However, even for midsized circuits, when the number of coupling nodes is beyond 300400, efficient iterative solvers become more competitive than direct solvers.
Our approach is based on the ILU(Ïµ)preconditioned GMRES Krylov space algorithm, cf. [34â€“36]. This iterative solver was already successfully used inÂ [24] as an efficient coupling system solver. While the author used a constant \(\epsilon=0.001\), we extend this approach by using an adaptive Ïµstrategy. Previously inÂ [24] this approach was tested for one constant matrix. During the transient simulation using the same threshold value Ïµ, as activity might change in the circuit, the pattern of the ILU(Ïµ) preconditioner might also change significantly. On the other hand, recomputing the pattern of the ILU(Ïµ) preconditioner poses a significant computational overhead. Therefore the convergence of the preconditioned GMRES and the computational overhead for an ILU(Ïµ) update needs to be balanced. If the pattern stays the same, then one simple measure is to only update the ILUdecomposition of the matrix, considering the current values of the Newton iteration matrix.
In summary, we developed an adaptive Ïµstrategy for our ILU preconditioner which detects poor convergence or divergence of the GMRES by monitoring the number of required iterations. Depending on the level of divergence one of the following steps is being executed:

1.
Update the ILUdecomposition using the current pattern and values of A.

2.
Recompute the pattern of ILU(Ïµ) with the value of Ïµ and do (1).

3.
Decrease Ïµ by two and do (2).
We use a starting value of \(\epsilon=0.05\) which for average simulations remains unchanged or decreases only slightly, so that the ILU(Ïµ) preconditioner remains computationally cheap. The values of Ïµ are limited by a lower bound of 10^{âˆ’5}, and in case of divergence of the iterative solver for \(\epsilon\le10^{5}\) a direct solver is used instead. With this backup strategy we do not just ensure robustness of transient simulations but we also avoid too small values of Ïµ.
One additional important aspect is that the core GMRES method consists mainly of matrix vector multiplicationsÂ [35] which can be parallelized efficiently, but the ILU preconditioner, for such matrix sizes, does not run efficiently in parallel. Therefore, large values of Ïµ favor parallel simulation since the computations in ILU(Ïµ) are marginal compared to GMRES. On the other hand, with small values of Ïµ a more accurate preconditioner is created and fewer GMRES iterations might be required. For the enlisted reasons it is crucial to automatically select the optimal values of Ïµ for a given transient simulation and circuit, extracting the maximal efficiency for sequential and parallel transient simulations.
2.4 Partition bypass acceleration
In the previous two sections, we presented the main components of our approach. The partitioner, the parallel matrix building and solving method result already in substantial transient simulation speedups. In this section, we present one additional acceleration method that exploits the latency in the partitioned circuit while maintaining the singlerate and SPICEaccuracy aspects of the transient simulation. Furthermore, this acceleration technique is built on top of the previous methods and can be deactivated.
The characteristic of large and midsized circuits is that during their operation mode the activity is mainly concentrated within a few number of partitionsÂ [10, 11]. In this analog context, â€˜activityâ€™ means the process when a semiconductor changes its state from ON to OFF or vice versa. During this transition the semiconductors have highly nonlinear behavior and SPICElevel accuracy is crucial to capture this transition. Furthermore, such nonlinear transitions can trigger a chain of other transitions at the same time during the Newton loop. Therefore it is crucial for SPICElevel accuracy to simulate the partitions in a singlerate way.
The main idea of the partition bypass is to reuse the factorized \(A_{i,i}\), the righthand side \(r_{i}\), and the contributions to the coupling system \(S_{i}\) and \(s_{i}\) from the previous Newton iteration, if the bypass criterions are met. Since we operate directly with the partition matrix \(A_{i,i}\) and with the righthand side \(r_{i}\), as the Newton linearization (5) shows, we cannot reuse the matrices from previous time steps, while the integration coefficientÂ Î± changes with the time step size and integration method. Therefore the bypass method starts with the second Newton iteration, such that \(A_{i}\), \(r_{i}\), \(S_{i}\) and \(s_{i}\) are computed with the correct Î±. Hence, each device is evaluated at least once in a time step ensuring the singlerate aspect of our approach.
FigureÂ 4 presents the modified Newton loop of FigureÂ 3. The difference to FigureÂ 3 is only in step (4.1). If for a partition i the bypass criterion is fulfilled, then the whole computation is skipped and the results from the previous Newton loop are reused. Steps (4.2) and (4.3) are computed as in FigureÂ 3. If a partition is bypassed for a Newton iteration, this does not imply that for subsequent Newton iterations it will be skipped as well, since in step (4.2) \(x_{c}\) is updated and in step (4.3) even a bypassed partitionâ€™s unknowns \(x_{i}\) will be changed.
As next we introduce the partition bypass criterion that is the key for the success of this acceleration. A strict partition bypass criterion increases the speedup only marginally, whereas inaccurate criterions can cause convergence problems or can even produce wrong results in certain cases.
We denote by \(\hat{x}_{i}\) the values of the unknowns \(x_{i}\) of partition i where the last \(A_{i,i}\) and \(r_{i}\) was built. Since not all of the elements \(x_{c}\) contribute to all partitions, we denote by \(x_{c}^{(i)}\) the subset of \(x_{c}\) that contributes to partition i, and \(\hat{x}_{c}^{(i)}\) denotes the last values of \(x_{c}^{(i)}\) where the partition i was evaluated. Newton convergence of partition i is monitored by the vector \(T_{i}\) which has the values
Here \(\mathrm{imax}_{i;j}\) is the absolute value of the maximal current contribution to voltage node j within partition i, and Pi represents the MNA unknowns \(x_{i}\) and those MNA unknowns from \(x_{c}\) that contribute to partition i. The absolute tolerance ABSTOL is set to 10^{âˆ’12} for the voltage nodes and to 10^{âˆ’6} for the current equations. The relative tolerance RELTOL is 10^{âˆ’3} for all MNA unknowns. The indicator vector represents the convergence state of the MNA variables during the Newton loop. The righthandside vector is the sum of the currents at a given node, and this sum, according to Kirchhoffâ€™s current law, should converge numerically to zero. If \(T_{i;j} \le1.0\), then the jth unknown of partition i is converged in terms of the Newton criterion. The Newton iteration is converged for partition i when \(\T_{i}\_{\infty} \le1.0\). With the introduced notation we define the bypass indicator for all MNA variables belonging to a partition \(i=1,\ldots,p\):
where \(\epsilon_{x}=10^{6}\) and \(\epsilon_{N}=0.8\). The tolerance \(\epsilon_{x}\) represents the trust region of the last evaluation point \(\hat{x}_{i}\) and \(\hat{x}_{c}^{(i)}\) where partition i was evaluated. The coefficient \(\epsilon_{N} \le1.0 \) ensures that the partition is converged in terms of the Newton criterions. The indicator \(B_{i}\) in (15) has a value of 1 when the new evaluation point for partition i is within the trust region of the last evaluation point or when the Newton is already converged for partition i. If the value of the indicator function \(B_{i}\) is zero, it means that the partition i should not be bypassed.
The bypass indicators (15) are evaluated at the beginning of each Newton iteration. The computational overhead for this block bypassing consists in storing the partitionâ€™s LUfactorized matrix and the righthand side, and in computing the bypass indicatorsÂ (15). Thus it does not pose significant additional computations compared to FigureÂ 3.
The bypass method of FigureÂ 4 transforms the Newton method into a QuasiNewton method by using a reduced model for the partition. If a partition is bypassed, then a constant extrapolation is used with the values of the last evaluation point. The possibility of linear extrapolation has also been studied inÂ [37], but it turned out that the constant extrapolation gives the best costbenefit ratio overall for robust transient simulations.
2.5 Applicability of the BBD solver for other analyses
So far we investigated the concept and application of our BBD matrix and solver approach in the context of transient analyses. In the following we emphasize some of the highlevel aspects of our approach applied to other circuit analyses that are often used beside transient analysis. The basic principle that each device contributes additively to the overall system is still true for DC and AC analyses. Therefore our concept of partitioning the circuit and solving the respective BBD system can be correspondingly used also for these analyses.
Apart from the absence of dynamic contributions, each step of a DC analysis can be considered as computing one transient timestep. Therefore the solving approach of the whole BBD system is the same as for transient analyses, and it might result in substantial performance gains compared to conventional solvers.
A smallsignal AC or AC NOISE analysis is built upon a linearization around an operating point and results in a sequence of linear problems for the selected frequency points. Then the whole BBD system becomes complexvalued, but is still structurally equivalent to the respective transient system.
In both cases we solve the coupling system of the BBD matrix with a direct solver, allowing to use the same matrix structure and solver for all these analyses. Further detailed elaboration of our matrix and solver approach in the context of AC and DC analysis is beyond the scope of this paper.
3 Results and discussion
In this section, we demonstrate the speedup potential of our presented method to significantly speed up transient analog circuit simulations. The implementation of the presented approach was done in Infineonâ€™s inhouse SPICE simulator TITAN, and all the numerical comparisons were made in the frame of this simulator, assuring that the numerical methods are tested in the same environment. The TITAN simulator is used in productive environment and its implementation is trimmed for high performance computing, therefore the presented novel method is tested in practical productive environment, showing the true potential of our approach.
For the numerical comparison we consider the nine circuits from TableÂ 1. These circuits represent a wide range of circuit sizes and types. We consider the semiconductordominated circuits cir1, cir2, cir3, cir6, cir7, and cir8, whereas the RCdominated circuits are cir4, cir5, and cir9. Furthermore we consider the midsized circuits where the total number of MNA variables is less than 10^{6}, these circuits are cir1, â€¦, cir5, whereas the large circuits are cir6, â€¦, cir9. As shown in TableÂ 1, the number of nonzero entries in the system matrix is increasing with the increasing number of MNA unknowns.
In an industrial context the RCdominated circuits are always preprocessed by a stateoftheart SPICEin, SPICEout network reduction toolÂ [18, 19]. This is also the case for the input circuits in our test suite. Hence the speedup of such network reduction techniques is orthogonal to the speedup of actual SPICElike simulations, and not taken into account in the following comparison.
As a base line of comparison we choose the KLU solverÂ [23, 38] within the TITAN simulator. This solver is well suited for sparse matrices that arise from analog circuit simulation, and this solver is also widely used in opensourceÂ [4, 5] and commercial SPICE simulators. Since KLUÂ [23, 38] does not have a parallel version, we consider the elapsed time of the sequential simulation as a reference for the sequential and parallel simulations with our approach.
The measured elapsed time results of the sequential and parallel simulations are presented in TableÂ 2. The number of partitions and the bypass rate of the partitions are presented in TableÂ 3. We also underline here that the presented circuits and their setup are taken from an industrial and practical context. Therefore they contain output and other sequential parts that by Amdahlâ€™s law limits the theoretical speedup of the total simulation time. For these reasons, we only scale up to 8 CPUs and we aim at a speedup in a range of \([3,4]\) with 8 CPUs compared to the same sequential run.
In the first step we consider the midsized circuits, cir1, â€¦, cir5. Within this group there are both RC and semiconductordominated circuits. cir1, cir2, and cir3 are semiconductordominated and our approach matches the sequential performance of the KLU solver. In these cases the partition bypass strategy also significantly improves the BBD solver performance. Especially for cir1 and cir3 we get a partition bypass rate around 30% (see TableÂ 3) and thus additional overall sequential speedup of 20% in the elapsed time. The parallel scaling for these circuits is also satisfactory since the speedup is around factor 3 with 8 CPUs.
However for RC dominated circuits, cir45, the parallel scaling only reaches a factor of 2 with 8 CPUs. For such circuits, as TableÂ 3 shows, the coupling system is relatively large compared to the overall system matrix, and the presented ILU preconditioned GMRES solver for the coupling system is not well suited for parallelization. Furthermore, the evaluation of simple devices such as linear resistors and capacitors represents a significantly smaller and parallelizable computation task than complex semiconductor models. For these reasons, the speedup factor compared to the semiconductordominated circuits is significantly lower. In comparison to the KLU solver, our BBD solver performs significantly better for cir5, but for cir4, due to the small circuit size, KLU performs sequentially 20% better.
The largesized circuits truly display the true potential of our approach. For the semiconductordominated circuits cir6 and cir7 the speedup compared to KLU is substantial, and additionally we get more than factor 4 speedup with 8 CPUs. Overall, in comparison with the sequential KLU solver in both cases we get double digit factor speedup factors, which yield a huge step forward in the analog transient circuit simulation.
For the RCdominated large circuits the sequential speedup of the BBD compared to the KLU is substantial and for very large circuits can have double digits. On the other hand in these cases the parallel speedup of the BBD solver is only around factor 2. This bad behavior is due to the ILU preconditioner which runs sequentially at the moment. For such very large circuits this limits the parallel scalability of our approach.
Another important aspect is the effect of the partition bypassing. As it is shown in TableÂ 2, turning on the partition bypassing in sequential mode does not have any significant effect on the number of time steps nor on the number of iterations. Therefore the presented approach for partition bypass represents a robust method to speed up the presented BBD solver. TableÂ 3 summarizes the partition bypass rate which depends not just on the size of the circuit or number of partitions, but mostly on the scenario that is simulated. A circuit startup scenario usually results in single digit partition bypass rates, since the supply voltage ramp up usually affects the whole circuit and results in activity in all partitions. On the other hand for circuits in the normal working regime at a given time point the activity is concentrated in a relatively small portion of the circuit.
In this section we demonstrated that the presented approach for transient analog circuit simulation is capable of double digit speedups for largesized circuits in comparison to the stateoftheart KLU solver.
4 Conclusion and outlook
In this publication we presented a holistic approach to achieve double digit speedups for analog transient simulation of largesized circuits compared to existing stateoftheart KLUÂ [23, 38] solver. The novelty of our approach consists in the combination and extension of existing approaches in a unique and unprecedented way. In the first step of our approach we partition the circuit such that the system matrix and righthand side can be built in parallel, while minimizing the fillin rate in the resulting coupling system of the BBD matrix. For solving the coupling system during the simulation we introduced an adaptive Ïµ approach for the ILU(Ïµ) preconditioned GMRES solver which solves the coupling part efficiently. As an additional speedup measure we introduced the partition bypass method. If certain criterions are met during the Newton iteration of a time step, the partition bypass method skips significant computations of the Newton loop. The numerical examples clearly underline not just the robustness of our approach but its true speedup potential especially for largesized circuits in the frame of an industrial analog simulator.
Further work should be focused on the ILU(Ïµ) preconditioner of the GMRES solver which represents a performance bottleneck for largesized circuits. Furthermore the GMRES solver should be coupled to the Newton convergence criterion such that fewer GMRES iterations are computed. The current form of the partition bypassing is also rather simple and could be also further extended to increase the bypass ratio of the partitions.
Notes
Netlist is the input file format that describes the hierarchical structure of the circuit.
\(10^{4} \leq n \leq10^{6}\).
Circuit where the linear resistors and capacitance dominate the circuit. This type of circuit is also called postlayout.
The number of MOSFETs, bipolar transistors, diodes, and other semiconductors are dominant in the circuit. This is also called prelayout circuit.
\(n \le10^{3}\).
\(10^{3} < n \le 10^{6} \).
\(n > 10^{6} \).
For the sake of simplicity we use \(A=A(x_{i})\).
References
Nagel LW, Pederson DO. SPICE (simulation program with integrated circuit emphasis). Berkeley: EECS Department, University of California; 1973. Technical Report UCB/ERL M382. http://www2.eecs.berkeley.edu/Pubs/TechRpts/1973/22871.html.
Cadence Design Systems, Inc. Spectre user manual. Cadence Design Systems, Inc. 2017. https://www.cadence.com/content/cadencewww/global/en_US/home/tools/customicanalogrfdesign/circuitsimulation/spectrecircuitsimulator.html.
Synopsys, Inc. HSPICE user manual. Synopsys, Inc. 2017. https://www.synopsys.com/verification/amsverification/circuitsimulation/hspice.html.
Nenzi P, Vogt H. Ngspice user manual. 2014. http://ngspice.sourceforge.net/docs/ngspice26manual.pdf.
Keiter ER, Aadithya KV, Mei T, Russo TV, Schiek RL, Sholander PE, Thornquist HK, Verley JC. Xyce parallel electronic simulator reference guide. 2016. https://xyce.sandia.gov/downloads/_assets/documents/Reference_Guide.pdf.
FrÃ¶hlich N, Riess BM, Wever UA, Zheng Q. A new approach for parallel simulation of VLSI circuits on a transistor level. IEEE Trans Circuits Syst I, Fundam Theory Appl. 1998;45(6):60113.
Feldmann U, Schultz R. TITAN: a universal circuit simulator with event control for latency exploitation. In: Fourteenth European solidstate circuits conference, ESSCIRC 88. 1988. p. 1835.
Feldmann U, Wever UA, Zheng Q, Schultz R, Wriedt H. Algorithms for modern circuit simulation. Arch Elektron Ãœbertragtech. 1992;46(4):27485.
GÃ¼nther M, Feldmann U, ter Maten J. Modelling and discretization of circuit problems. In: Schilders WHA, ter Maten EJW, editors. Handbook of numerical analysis, vol. XIII. Special volume: numerical methods in electromagnetics. Amsterdam: Elsevier; 2005. p. 523659.
Rewienski M. A perspective on fastSPICE simulation technology. Simulation and verification of electronic and biological systems. Berlin: Springer; 2011. doi:10.1007/97894007014962.
Li P. Parallel circuit simulation: a historical perspective and recent developments. Found Trends Electron Des Autom. 2012;5(4):211318. doi:10.1561/1000000020.
Deng AC. On network partitioning algorithm of largescale CMOS circuits. IEEE Trans Circuits Syst. 1989;36:2949. doi:10.1109/31.20209.
Deng AC, Tuan JF, Ong LW. An investigation on parasitic couplings and feedback loops in the CMOS circuits. In: 1989 IEEE international symposium on circuits and systems. Vol.Â 2. 1989. p. 8647. doi:10.1109/ISCAS.1989.100488.
Peng H, Cheng CK. Parallel transistor level circuit simulation using domain decomposition methods. In: Proceedings of the 2009 Asia and South Pacific design automation conference, ASPDAC â€™09. Piscataway: IEEE Press; 2009. p.Â 397402. http://dl.acm.org/citation.cfm?id=1509633.1509732.
Tcherniaev A, Feinberg I, Chan W, Tuan JF, Deng AC. Transistor level circuit simulator using hierarchical data. Google Patents. 2003. US Patent 6,577,992. https://www.google.es/patents/US6577992.
Lin PS, Wen KS, Perng RK. Simulation of circuits with repetitive elements. Google Patents. 2014. US Patent 8,832,635. https://www.google.com/patents/US8832635.
Striebel M. Hierarchical mixed multirating for distributed integration of DAE network equations in chip design [PhD thesis]. University of Wuppertal, Department of Applied Mathematics and Numerical Analysis; 2006.
Kerns KJ, Yang AT. Stable and efficient reduction of large, multiport RC networks by pole analysis via congruence transformations. IEEE Trans ComputAided Des Integr Circuits Syst. 1997;16:73444.
Ionutiu R, Rommes J, Schilders WHA. SparseRC: sparsity preserving model reduction for RC circuits with many terminals. IEEE Trans ComputAided Des Integr Circuits Syst. 2011;30:182841.
Chen X, Wang Y, Yang H. NICSLU: an adaptive sparse matrix solver for parallel circuit simulation. IEEE Trans ComputAided Des Integr Circuits Syst. 2013;32(2):26174.
Chen X, Wu W, Wang Y, Yu H, Yang H. An ESchedulerbased data dependence analysis and task scheduling for parallel circuit simulation. IEEE Trans Circuits Syst II, Express Briefs. 2011;58(10):7026. doi:10.1109/TCSII.2011.2164148.
Chen X, Ren L, Wang Y, Yang H. GPUaccelerated sparse LU factorization for circuit simulation with performance modeling. IEEE Trans Parallel Distrib Syst. 2015;26(3):78695.
Davis TA, Palamadai Natarajan E. Algorithm 907: KLU, a direct sparse solver for circuit simulation problems. ACM Trans Math Softw. 2010;37(3):Article No. 36. doi:10.1145/1824801.1824814.
Bomhof CW. Iterative and parallel methods for linear systems with application in circuit simulation [PhD thesis]. University Utrecht, Computer Science Department; 2001.
Kapre N, DeHon A. Accelerating SPICE modelevaluation using FPGAs. In: 2009 17th IEEE symposium on field programmable custom computing machines, FCCM â€™09. 2009. doi:10.1109/FCCM.2009.14.
Bayoumi AM, Hanafy YY. Massive parallelization of SPICE device model evaluation on GPUbased SIMD architectures. In: Proceedings of the 1st international forum on nextgeneration multicore/manycore technologies, IFMT â€™08. New York: ACM; 2008. Article No. 12. doi:10.1145/1463768.1463784.
Vlach M. Modeling and simulation with Saber. In: Proceedings of the third annual IEEE ASIC seminar and exhibit. 1990. doi:10.1109/ASIC.1990.186080.
Yu H, Chu C, Shi Y, Smart D, He L, Tan SXD. Fast analysis of a largescale inductive interconnect by blockstructurepreserved macromodeling. IEEE Trans Very Large Scale Integr (VLSI) Syst. 2010;18(10):1399411.
Miettinen P, Honkala M, Roos J. Using METIS and hMETIS algorithms in circuit partitioning. 2006.
Karypis G, Kumar V. METIS  unstructured graph partitioning and sparse matrix ordering system, version 2.0. Minneapolis: Department of Computer Science, University of Minnesota; 1995. Technical report.
Chevalier C, Pellegrini F. PTScotch: a tool for efficient parallel graph ordering. Parallel Comput. 2008;34(68):31831. doi:10.1016/j.parco.2007.12.001.
Devine KD, Boman EG, Riesen LA, Catalyurek UV, Chevalier C. Getting started with Zoltan: a short tutorial. In: Proceedings of the 2009 Dagstuhl seminar on combinatorial scientific computing. 2009. Also available as Sandia National Labs Technical Report SAND20090578C.
Basermann A, CortialGoutaudier F, Jaekel U, Hachiya K. Parallel solution techniques for sparse linear systems in circuit simulation. Math Ind. 2004;4:1129. doi:10.1007/9783642558726_10.
Saad Y. Iterative methods for sparse linear systems. 2nd ed. Philadelphia: Society for Industrial and Applied Mathematics; 2003.
Saad Y, Schultz MH. GMRES: a generalized minimal residual algorithm for solving nonsymmetric linear systems. SIAM J Sci Stat Comput. 1986;7(3):85669. doi:10.1137/0907058.
Davis TA. Direct methods for sparse linear systems. Philadelphia: Society for Industrial and Applied Mathematics; 2006. doi:10.1137/1.9780898718881.bm.
Syed AA. Fast quasiNewton method for partitioned analog circuit simulation in time domain [masterâ€™s thesis]. Technische UniversitÃ¤t MÃ¼nchen, Computer Science Department; 2016.
Davis TA, Natarajan EP. Sparse matrix methods for circuit simulation problems. Math Ind. 2012;16:314. doi:10.1007/9783642224539_1.
Acknowledgements
We acknowledge the support of Infineon Technologies AG for this publication.
Funding
This research was completely supported by Infineon Technologies AG.
Author information
Authors and Affiliations
Contributions
The main idea of this paper was proposed by JB and the implementation was done also by JB. GD and KW significantly supported JB in the implementation and helped preparing the manuscript. All authors read and approved the final manuscript.
Corresponding author
Ethics declarations
Competing interests
The authors declare that they have no competing interests.
Additional information
Publisherâ€™s Note
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
Rights and permissions
Open Access This article is distributed under the terms of the Creative Commons Attribution 4.0 International License (http://creativecommons.org/licenses/by/4.0/), which permits unrestricted use, distribution, and reproduction in any medium, provided you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons license, and indicate if changes were made.
About this article
Cite this article
Benk, J., Denk, G. & Waldherr, K. A holistic fast and parallel approach for accurate transient simulations of analog circuits. J.Math.Industry 7, 12 (2017). https://doi.org/10.1186/s133620170042z
Received:
Accepted:
Published:
DOI: https://doi.org/10.1186/s133620170042z