4 Power converters\index{power converters} have seen a surge of new trends and novel
5 applications due to their widespread use in renewable energy systems
6 and emerging hybrid and purely-electric vehicles. More efficient
7 simulation techniques for power converters are urgently needed to meet
8 more design constraints. In this chapter, we present a novel
9 envelope-following parallel transient analysis method for the general
10 switching power converters. The new method first exploits the
11 parallelisim in the envelope-following method and parallelize the
12 Newton update solving part, which is the most computational expensive,
13 in GPU platforms to boost the simulation performance. To further
14 speed up the iterative GMRES\index{iterative method!GMRES} solving for Newton update equation in the
15 envelope-following method, we apply the matrix-free\index{matrix-free}
16 Krylov subspace\index{Krylov subspace} basis
17 generation technique, which was previously used for RF simulation.
18 Last, the new method also applies more robust Gear-2 integration to
19 compute the sensitivity matrix instead of traditional integration
22 Over the past decades, power electronics\index{power electronics} and especially switching
23 power converters have seen a surge of new trends and novel applications
24 --- from the growing significance of PWM (pulse-width modulation)
25 rectifiers and multilevel inverters to their widespread use in
26 renewable (like solar and wind) energy systems, smart grids and emerging
27 electric and hybrid vehicles~\cite{Trzynadlowski:Book'10}.
29 This requires more efficient simulation techniques for power
30 electronics to meet the new application and demanding design
31 constraints. To facilitate the design of typical power electronics
32 circuits, many special-purpose simulation algorithms and
33 tools were developed. Among them is the envelope-following
34 method~\cite{Kundert:ICCAD'88,White:TPE'91,Feldmann:ICCAD'96}, %,Bose:TIE'09
35 which is able to calculate the slowly changing contour, or
36 envelope, of a carrier waveform with a much higher
37 switching frequency\index{switching frequency}.
38 This is the case for switching power converters,
39 which have fast switching currents to convert powers
40 from one level to another level.
41 In those switching power converters, it is the envelope,
42 which is the power voltage delivered,
43 not the fast switching waves in every cycle,
44 that is of interest to the designers.
45 As shown in Fig.~\ref{fig:ef1}, the solid line is
46 the waveform of the output node in a Buck
47 converter~\cite{Krein:book'97}, the dots are the simulation points
48 of SPICE\index{SPICE}, and the appended dash line is the envelope.
50 Obviously, traditional SPICE will incur an extraordinarily high
51 simulation time for this task, since it has to integrate the
52 circuit's differential equation with many time points in every
53 clock cycle to get the accurate details of the carrier. For
54 switching power converters, the waveform of the carrier in
55 consequent cycles does not change much, envelope-following method
56 is an approximation analysis method, which skips over several
57 cycles (the dash line in Fig.~\ref{fig:ef2}), the so called
58 envelope step, without simulating them, and then carries out a
59 correction, which usually contains a sensitivity-based Newton
60 iteration or shooting until convergence, in order to begin the
64 % \resizebox{.5\textwidth}{!}{\input{./matlab/ef_intro.pstex_t}}
65 % %\includegraphics[width=.45\textwidth]{./matlab/ef_intro_ppt.eps}
66 % \caption{Transient envelope-following analysis.
67 % (This figure reflects a backward-Euler style envelope-following.)}
68 % \label{fig:ef_intro}
72 \subfigure[Illustration of one envelope skip.]
73 {\resizebox{.8\textwidth}{!}{\input{./Chapters/chapter16/figures/ef_intro.pdf_t}}
75 \subfigure[The envelope changes in a slow time scale.]
76 {\resizebox{.9\textwidth}{!}{\input{./Chapters/chapter16/figures/envelope.pdf_t}}
78 \caption{Transient envelope-following\index{envelope-following} analysis.
79 (Both two figures reflect backward Euler\index{backward Euler} style envelope-following.)}
85 Also, iterative GMRES\index{iterative method!GMRES} solver is typically used in the
86 envelope-following method to compute the solution of Newton update
87 due to its efficiency compared to direct LU\index{LU} method.
88 However, as the Jacobian matrix\index{Jacobian matrix}
89 or the sensitivity matrix\index{sensitivity matrix}
90 in the equation to be solved is dense,
91 explicit computing of the Jacobian is a very expense process.
92 Recently, the matrix-free\index{matrix-free} GMRES
93 was proposed~\cite{Telichevesky:DAC'95}
94 for RF shooting based simulation.
95 The new method leads to significant
96 savings due to its implicit calculation of new basis vectors
97 without the explicit formulation of the sensitivity matrix.
100 Modern computer architecture has shifted towards designs that employ
101 so called multi-core processor or chip-multiprocessors (CMP)~\cite{IntelMC'06,AMDMC'06}.
102 The family of graphic processing units (GPU) are among the most
103 powerful many-core computing systems in mass-market use~\cite{nvidia}.
104 For instance, the state-of-the-art NVIDIA Tesla T10 chip has a peak
105 performance of over 1~TFLOPS versus about 80--100~GFLOPS of Intel i5
106 series Quad-core CPUs~\cite{Kirk:Book'10}.
107 In addition to the primary
108 use of GPUs in accelerating graphics rendering operations, there has
109 been considerable interest in exploiting GPUs for general purpose
110 computation (GPGPU)~\cite{gpgpu}. The introduction of new parallel
111 programming interfaces for general purpose computation, such as
112 Computer Unified Device Architecture (CUDA),
113 Stream SDK, and OpenCL~\cite{cuda,streamSDK,openCL},
114 has made GPUs an attractive choice
115 for developing high-performance scientific computation tools
116 and solving practical engineering problems.
117 Hence, the applications with GPGPUs are rapidly growing in a broad
118 variety of parallel numerical computation works~\cite{CUDA_Zone}.
120 %For this correction to be efficient and precise, it is critical
121 %for the Newton iteration to converge fast and yield accurate
124 %Switching power converters and RF (radio-frequency) circuits are
125 %typically stiff systems where
126 %the fast signals (switching signals or carrier signals) and
127 %the slow moving signals (the output voltage waveforms and user signals)
129 %However, it is challenging to simulate the stiff systems
130 %using the traditional integration method such as
131 %Backward Euler and Trapezoidal method~\cite{Iserles:Book'96}.
132 %It has been shown that Gear-2 is more suitable
133 %for many practical problems such as stiff problems
134 %where one have different time constants
135 %(fast ones with large poles and slow ones with small poles)~\cite{Vlach:Book'94}.
137 %GMRES is a Krylov-subspace based method for solving linear
138 %equation, which guarantees convergence and is efficient if the
139 %eigenvalues are tightly clustered together~\cite{Golub:Book'96}.
140 %Fortunately, most problems arising from the shooting method and
141 %the envelope-following method satisfy naturally the convergence
142 %criterion of GMRES~\cite{Nastov:ProcIEEE'07} (or after
143 %preconditioners applied if not).
146 %\item Newton update equation for envelope-following is formed with
147 % the sensitivity matrix derived from the Gear-2 integration rule for
149 % \item Detailed mathematical derivations to show
150 % the relationship between the sensitivity matrix and
151 % the information extracted from transient steps.
153 After the basic algorithm of envelope-following is briefly reviewed
154 and Newton update equation is derived in Section~\ref{sec:ef}. In
155 Section~\ref{sec:gmres} presents the new parallel envelope method
156 where the CPU parallelization, matrix-free GMRES and Gear-2
157 integration will be discussed. We will show the parallelization of the
158 Newton update, which is the most time consuming step, in the
159 envelop-follow method, in GPU platforms. We will also present the new
160 GMRES solver using matrix-free Jacobian-vector multiplication and the
161 Gear-2 integration for sensitivity based Newton update equation.
162 Numerical examples are shown in Section~\ref{sec:exp}, and finally,
163 this chapter is summarized in Section~\ref{sec:summary}.