Asymptotic probability extraction for non-normal distributions of circuit performance
Summary (5 min read)
Introduction
- As IC technologies are scaled to the deep sub-micron region, process variations are becoming critical and significantly impact the overall performance of a circuit.
- Table 1 shows some typical process parameters and their 3 variations as technologies are scaled from 0.25 µm to 70 nm.
- These large-scale variations introduce uncertainties in circuit behavior, thereby making IC design increasingly difficult.
- Low product yield or unnecessary over-design cannot be avoided if these process variations are not accurately modeled and analyzed within the IC design flow.
2. Background
- Given a circuit topology, the circuit performance (e.g. gain, delay) is a function of the design parameters (e.g. bias current, transistor sizes) and the process parameters (e.g. VTH, TOX).
- Given a set of fixed design parameters, the circuit performance f can be approximated by a linear regression model [1]: XBXfXf Tˆ (1) where TNxxxX ,,, 21 denotes the process parameters, X is the mean value of X, XXX represents the process variations, NRB̂ stands for the linear model coefficients and N is the total number of these random variations.
- The process variations in (1), i.e. X, are often approximated by zero-mean Normal distributions*.
- Given a set of normally distributed random variables X and their symmetric, positive semi-definite correlation matrix R, PCA decomposes R as: TVVR (2) where Ndiag ,,, 21 contains the eigenvalues of R, and NVVVV ,,, 21 contains the corresponding eigenvectors that are orthonormal, i.e. IVV T (I is the identity matrix).
- Without loss of generality, the authors assume that A is symmetric in this paper, since any asymmetric quadratic form can be easily converted to an equivalent symmetric form [10].
3. Asymptotic Probability Extraction
- Given the quadratic response surface model in (5), the objective of probability extraction is to estimate the unknown probability density function fpdf and cumulative distribution function fcdf for performance f.
- Instead of running expensive Monte Carlo simulations, APEX tries to find an M-th order LTI system H whose impulse response th and step response ts are the optimal approximations for the fpdf and fcdf respectively.
- The optimal approximation is determined by matching the first 2M moments between th and fpdf for an M-th order approximation.
- The authors first describe the mathematical formulation of APEX in Section 3.1.
- Then, in Section 3.2 the authors will link APEX to traditional probability theory and explain why it can be used to efficiently approximate PDF/CDF functions.
3.1 Mathematical Formulation
- The authors define the time moments [11] for a given circuit performance f whose probability density function is fpdf as follows: dffpdff k m k k k !.
- 1 (6) In (6), the definition of time moments is identical to the traditional definition of moments in probability theory except for the scaling factor !.
- It is worth noting that many implementation issues must be considered to make their proposed approach, APEX, feasible and efficient.
- The impulse response of a causal LTI system is only nonzero for 0t , but a PDF in practical applications can be nonzero for 0f .
- In the following section, the authors will explain why this momentmatching approach is an efficient way to approximate PDF/CDF functions.
3.2 Connection to Probability Theory
- In probability theory, given a random variable f whose probability density function is fpdf , the characteristic function is defined as the Fourier transform of fpdf [8].
- This implies an important fact: the time moments defined in (6) are related to the Taylor expansion of the characteristic function at the expansion point 0 .
- Therefore, the optimally approximated sH in (7) is a low pass system.
- It is well-known that a Taylor expansion is accurate around the expansion point.
- This conclusion has been verified in other applications (e.g. IC interconnect order reduction [11], [12]) and it provides the theoretical background to explain why moment-matching works well for the PDF/CDF evaluations that the authors will demonstrate in Section 5.
4. Implementation of APEX
- The proposed APEX approach is made practically feasible by applying several novel algorithms, including: 1) a binomial scheme for high order moment computation, 2) a generalized Chebyshev inequality for PDF/CDF shifting and 3) a reverse evaluation technique for best-case/worst-case analysis.
- The authors describe the mathematical formulation of each of these algorithms.
4.1 Binomial Moment Evaluation
- A key operation required in APEX is the computation of the high order time moments defined in (6) for a given random variable f.
- Given the quadratic response surface model in (5), kf is a high order polynomial in Y: i Ni k Niii yyycYf 21 21 (13) where iy is the i-th element in the vector Y, ic is the coefficient of the i-th product term and ij is the positive integer exponent.
- Since the random variables Y are independent after PCA analysis, the authors have: i Ni k Niii yEyEyEcfE 21 21 (14) where E stands for the expectation.
- The above computation scheme is called direct moment evaluation in this paper.
- The key disadvantage of the direct moment evaluation is that, as k increases, the total number of the product terms in (14) will increase exponentially, thereby quickly making their computation infeasible.
A. Quadratic Model Diagonalization
- The first step in binomial moment evaluation is to remove the cross product terms in the quadratic response surface model (5), thereby yielding a much simpler, but equivalent, quadratic model.
- According to matrix theory [10], any symmetric matrix NNRA can be diagonalized as: TUUA (16) where Ndiag ,,, 21 contains the eigenvalues of A and NUUUU ,,, 21 is an orthogonal matrix (i.e. IUU T ) containing the eigenvectors.
- Equation (18) implies that there is no cross product term in the quadratic model after diagonalization.
- Given a set of independent random variables Y with the Normal distribution 1,0N and an orthogonal matrix U, the random variables Z defined in (17) are independent and satisfy the Normal distribution 1,0N, also known as Theorem 1.
- The correlation matrix for Z is given by: UYYEUUYYUEZZE TTTTT (19) Remember that Y is a set of independent random variables with a Normal distribution 1,0N , i.e. IYYE T , and matrix U is orthogonal, i.e. IUU T .
B. Moment Evaluation
- Table 2 compares the computation time for direct moment evaluation and their proposed binomial moment evaluation.
- In direct moment evaluation, the number of the total product terms increases exponentially, thereby making the computation task quickly infeasible.
- Binomial moment evaluation, however, is extremely fast and achieves more than 106x speedup over direct moment evaluation.
- In addition, the authors verify that the moment values obtained from both approaches are identical except for numerical errors.
4.2 PDF/CDF Shifting
- APEX approximates the unknown PDF fpdf as the impulse response th of an LTI system.
- In such cases, the authors need to right-shift the unknown fpdf by 0f and use the impulse response th to approximate the shifted PDF 0ffpdf , as shown in Fig. 3 (case 1).
- The minimal value of all these k values is utilized as the final for PDF/CDF shifting, since the authors aim to find the smallest to achieve high approximation accuracy for fpdf .
- The authors have generalized the 2nd order Chebyshev inequality to higher orders and, therefore refer to (24) as the generalized Chebyshev inequality.
- In practical applications the authors find that high order moments provide a much tighter (i.e. smaller) estimation of , as is demonstrated by the numerical examples in Section 5.
4.3 Reverse PDF/CDF Evaluation
- In many practical applications, such as robust circuit optimization [4], [14], the best-case performance (e.g. the 1% point on CDF) and the worst-case performance (e.g. the 99% point on CDF) are two important metrics to be evaluated.
- As discussed in Section 3.2, APEX matches the first 2M Taylor expansion coefficients between the original characteristic function and the approximated rational function.
- Remember that the Taylor expansion is most accurate around the expansion point 0 .
- The above analysis motivates us to apply a reverse evaluation scheme for accurately estimating the 1% point.
- As shown in Fig. 4, the reverse evaluation algorithm flips the original fpdf to fpdf .
4.4 Summary
- Fig. 5 summarizes the overall implementation of APEX except for reverse evaluation.
- Kkk fEfE 1 (30) where the high order expectations kfE have already been calculated in previous computations, also known as Note that.
- These methods can also be applied here for APEX.
- In addition, it is worth mentioning that using an approximation order greater than 10 can result in serious numerical problems [11], [12].
- In most practical applications, the authors find that selecting M in the range of 7~10 can achieve the best accuracy.
5. Numerical Examples
- In this section the authors demonstrate the efficacy of APEX using several circuit examples.
- All experiments are run on a SUN Sparc 1GHz server.
A. Response Surface Modeling
- As a second example the authors consider a low noise amplifier designed in the IBM BiCMOS 0.25 µm process, as shown in Fig. 9.
- The variations on both MOS transistors and passive components (capacitance and inductance) are considered.
- The probability distributions and the correlation information of these variations are provided in the IBM design kit.
- The authors approximate these unknown functions by linear regression models * Table 4 shows the modeling error for all these 8 performances.
C. PDF/CDF Shifting
- As discussed in Section 4.2, PDF/CDF shifting is necessary to make the proposed APEX approach feasible and efficient.
- A key operation for PDF/CDF shifting is determining the value based on (28) (also see Fig. 3).
- Fig. 7 shows the estimated value using various high order moments.
- From Fig. 7, the authors find that the high order moments 2k provide a much tighter (i.e. smaller) estimation of .
- After the moment order 10k , further increases in k do not have a significant impact on reducing .
D. PDF/CDF Evaluation
- Fig. 8 shows the cumulative distribution function using various approximation orders.
- In Fig. 8, the “exact” cumulative distribution function is evaluated by Monte Carlo simulation with 106 samples.
- Note that, the CDF obtained from the low order approximation (Order = 4) is not accurate and contains numerical oscillations.
- Once the approximation order is increased to 8, these oscillations are eliminated and the approximated CDF asymptotically approaches the exact CDF.
- Similar behavior has been noted in moment matching of LTI models of interconnect circuits [11], [12].
E. Comparison of Accuracy and Speed
- Table 3 compares the accuracy and speed for three different probability extraction approaches: linear regression, Monte Carlo analysis with 104 samples, and the proposed APEX approach.
- The 1% point and the 99% point, for example, denote the best-case delay and the worst-case delay respectively.
- The error values in Table 3 are calculated against the “exact” CDF obtained by Monte Carlo simulation with 106 samples.
- APEX achieves more than 200x speedup over the Monte Carlo analysis with 104 samples, while still providing better accuracy.
- This observation demonstrates the efficacy of the reverse evaluation method proposed in Section 4.3.
6. Conclusion
- As IC technologies reach nanoscale, process variations are becoming relatively large and nonlinear response surface models might be required to accurately characterize the large-scale variations.
- In this paper the authors propose an asymptotic probability extraction (APEX) method for estimating the nonNormal random distribution resulting from the nonlinear response surface modeling.
- Three novel algorithms, i.e. binomial moment evaluation, CDF/PDF shifting and reverse PDF/CDF evaluation, are proposed to reduce the computation cost and improve the estimation accuracy.
- As is demonstrated by the numerical examples, applying APEX results in better accuracy than the Monte Carlo analysis with 104 samples, and achieves more than 200x speedup.
- The efficacy of applying APEX to robust analog design is further discussed in [14].
Did you find this useful? Give us your feedback
Citations
344 citations
341 citations
234 citations
227 citations
Cites methods from "Asymptotic probability extraction f..."
...In combination with efficient numerical techniques like APEX [ 103 ] to estimate the pdfs of nonlinear circuits, ROAD can be used as an effective post-nominal robust optimizer....
[...]
197 citations
References
[...]
34,729 citations
"Asymptotic probability extraction f..." refers background in this paper
...[10] G....
[...]
...Without loss of generality, we assume that A is symmetric in this paper, since any asymmetric quadratic form can be easily converted to an equivalent symmetric form [10]....
[...]
...According to matrix theory [10], any symmetric matrix N N R A can be diagonalized as: T U U A (16) where N diag , , , 2 1 contains the eigenvalues of A...
[...]
4,897 citations
1,800 citations
"Asymptotic probability extraction f..." refers background or methods or result in this paper
...Similar behavior has been noted in moment matching of LTI models of interconnect circuits [11], [12]....
[...]
...The time moments of H are expressed as [11]:...
[...]
...Similarly, time moments can be defined for an LTI system H [11]....
[...]
...In addition, it is worth mentioning that using an approximation order greater than 10 can result in serious numerical problems [11], [12]....
[...]
...[11] L....
[...]
448 citations
"Asymptotic probability extraction f..." refers background or methods in this paper
...Technology parameters and 3 variations [1]...
[...]
...Given a set of fixed design parameters, the circuit performance f can be approximated by a linear regression model [1]:...
[...]
...The author in [1] applies linear regression to approximate a given circuit performance f (e....
[...]
...Reference [1] S....
[...]
...During the past decade, various statistical analysis techniques [1]-[7] have been proposed and utilized in many applications such as statistical timing analysis, mismatch analysis, yield optimization, etc....
[...]
204 citations
"Asymptotic probability extraction f..." refers background or methods or result in this paper
...In addition, it is worth mentioning that using an approximation order greater than 10 can result in serious numerical problems [11], [12]....
[...]
...The authors in [11] and [12] proposed several methods for iteratively determining M based on the approximation error....
[...]
...The aforementioned moment-matching method was previously applied for IC interconnect order reduction [11], [12] and is related to the Padé approximation in linear control theory [13]....
[...]
...[12] M....
[...]
...IC interconnect order reduction [11], [12]) and it provides the theoretical background to explain why moment-matching works well for the PDF/CDF evaluations that we will demonstrate in Section 5....
[...]
Related Papers (5)
Frequently Asked Questions (13)
Q2. How can a low-pass system be accurately approximated?
Since a low-pass system is mainly determined by its behavior in the low-frequency band (around 0 ), it can be accurately approximated by matching the first several Taylor coefficients at 0 , i.e. the moments.
Q3. What is the property that 1lh and lg are independent?
Equation (23) utilizes the property that 1lh and lg are independent, because 1lh is a function of1,,2,1, lizi , lg is a function of lz and all iz aremutually independent.
Q4. What is the way to evaluate the moment?
It should be noted that as long as the circuit performance f is represented by the quadratic model in (5) and the process variations are normally distributed, binomial moment evaluation provides the exact high order moment values (except for numerical errors).
Q5. What is the objective of the techniques?
The objective of these techniques is to model the probability distribution of the circuit performance under random process variations.
Q6. How can the authors determine the expectation of kf?
In addition, remember that each random variable iy has a Normal distribution 1,0N , which yields [8]:,6,4,2131,5,3,10 01 kk k k yE ki (15)Substituting (15) into (14), the expectation of kf can be determined.
Q7. What is the simplest way to approximate the circuit performance?
Given a set of fixed design parameters, the circuit performance f can be approximated by a linear regression model [1]: XBXfXf Tˆ (1)where TNxxxX ,,, 21 denotes the process parameters, Xis the mean value of X, XXX represents the process variations, NRB̂ stands for the linear model coefficients and N is the total number of these random variations.
Q8. How can the expectation iklzE be evaluated?
In (22), the expectation iklzE 2 can be easily evaluatedusing the closed-form expression (15), since lz is normally distributed 1,0N .
Q9. What is the best-case performance for fpdf?
In many practical applications, such as robust circuit optimization [4], [14], the best-case performance (e.g. the 1% point on CDF) and the worst-case performance (e.g. the 99% point on CDF) are two important metrics to be evaluated.
Q10. What is the key operation required by the binomial moment evaluation algorithm?
Instead of computing the high order moments of fdirectly, the proposed binomial moment evaluation scheme successively computes the moments of lh , as shown in Fig. 2.Step 2 in Fig. 2 is the key operation required by the binomial moment evaluation algorithm.
Q11. What is the normal distribution of the random variables in the quadratic model?
Theorem 1: Given a set of independent random variables Y with the Normal distribution 1,0N and an orthogonal matrix U, the random variables Z defined in (17) are independent and satisfy the Normal distribution 1,0N .
Q12. What is the correlation matrix for Z?
The correlation matrix for Z is given by: UYYEUUYYUEZZE TTTTT (19) Remember that Y is a set of independent random variables with a Normal distribution 1,0N , i.e. IYYE T , and matrix U isorthogonal, i.e. IUU T .
Q13. How can the authors guarantee high approximation accuracy?
the large-scale variations in the deep sub-micron technologies, which can reach35% as shown in Table 1, suggest applying high order regression models in order to guarantee high approximation accuracy [4]-[7].