layout: true class: typo, typo-selection --- count: false class: nord-dark, middle, center # ⚙️ Lecture 4: Robust Analog Circuit Sizing Under Process Variations @luk036 2025-10-22 📅 --- ## Abstract 📝 This lecture focuses on robust sizing of analog circuits under process variations at advanced nodes. Key topics include: - Challenges in 20nm analog design (double-patterning, layout effects, FinFET/GAA transistors) 🏗️ - Designing for robustness against process variations 🛡️ - Analog circuit sizing problem formulation ⚙️ - Emerging CAD approaches: - Knowledge-based methods - Optimization-based techniques 📈 - Robust geometric programming using: - Affine arithmetic ➕ - Ellipsoid method 🧮 - CMOS two-stage op-amp design case study 🔌 --- ### Keywords 🔑 .pull-left[ - Analog circuit 模拟电路 🎛️ - Design for robustness 鲁棒性设计 🛡️ - Worst-case scenarios 最坏情景 ⚠️ - Affine arithmetic 仿射运算 ➕ - Convex programming 凸规划 🥚 - Geometric programming 几何规划 📐 - Posynomial 正项式 (Positive + polynomial) ➗ - Ellipsoid method 椭球法 🫒 ] .pull-right[  ] --- ### Overview 🗺️ - 20nm Analog Design Challenges 🏗️ - Design for Variability 📏 - Design for Robustness 🛡️ - Analog Sizing Formulation ⚙️ - Robust Geometric Programming 🛡️ - Affine Arithmetic Methods ➕ - Design Examples 🔌 --- ### 🎬 Introduction: Cost Comparison | Costs | 28nm | 20nm | | ------------ | --------- | ----------- | | Fab Costs | 3B | 4B - 7B | | Process R&D | 1.2B | 2.1B - 3B | | Mask Costs | 2M - 3M | 5M - 8M | | Design Costs | 50M - 90M | 120M - 500M | : Fab, process, mask, and design costs are much higher at 20nm (IBS, May 2011) --- ### Challenges at 20nm 🏗️ - Double-patterning requirements 🎨 - Layout-dependent effects 📐 - New local interconnect layers 🔌 - \> 5,000 design rules 📜 - Device variation and sensitivity ⚠️ - FinFET/GAA transistors 🏗️ --- ### Double Patterning 🎨  --- ### Overlay Error (Mask Shift) ⚠️ - Parasitic matching becomes very challenging  --- ### Layout-Dependent Effects 📐 | Effect | >40nm | 40nm | ≥28nm | |---------------------------------|:-----:|:----:|:-----:| | Well Proximity Effect (WPE) | x | x | x | | Poly Spacing Effect (PSE) | | x | x | | Length of Diffusion (LOD) | x | x | x | | OD to OD Spacing Effect (OSE) | | x | x | --- ### New Local Interconnect Layers 🔌  --- ### FinFET Transistors 🏗️  --- ### Gate-All-Around Transistors 🏗️  --- ### Design for Robustness 🛡️ Key considerations for analog circuits: - Process variations are inherent in manufacturing - Variations caused by: - Temperature 🌡️ - Voltage ⚡ - Manufacturing tolerances 🏭 - Must include process variations in specifications --- ### Basic Design Flow 🔄  --- ### Top-down Design Phases 🔝  --- ### Analog Synthesis Flow 🔌  --- ### Analog Circuit Sizing Problem ⚙️ **Definition**: Given circuit topology, specifications and technology, find design variables that: - Meet all specifications - Optimize circuit performance **Challenges**: - High degrees of freedom 🎚️ - Performance sensitivity to variations ⚠️ --- ### Main CAD Approaches 💻 .pull-left[ **Optimization-based**: - Equation-based: - Circuit equations - Numerical solvers - Simulation-based: - Circuit simulation ] .pull-right[ **Knowledge-based**: - Circuit understanding - Design heuristics #### Practical designs mix these approaches ] --- ### 🎯 Robust Analog Sizing Flow .mermaid[ <pre> graph LR A[Input Specs] --> B[Choose Sizing] B --> C{Check Constraints} C -->|Fail| B C -->|Pass| D[Optimize Objective] D --> E[Robust Design] style A fill:#c8e6c9,stroke:#2e7d32,stroke-width:3px style B fill:#e3f2fd,stroke:#1565c0,stroke-width:3px style C fill:#fff3e0,stroke:#e65100,stroke-width:3px style D fill:#f3e5f5,stroke:#7b1fa2,stroke-width:3px style E fill:#c8e6c9,stroke:#2e7d32,stroke-width:3px </pre> ] --- count: false class: nord-light, middle, center ## Geometric Programming --- ### 📐 Geometric Programming (GP) Emerging technique for analog design: - Solves robust GP problems using: - Ellipsoid method 🫒 - Affine arithmetic ➕ - Provides mathematical guarantees - Enables formal optimization --- ### Circuit Sizing for Op Amp 🏗️ **Methodology**: 1. Geometric Programming formulation 2. Min-max convex programming under PVT variations 3. Ellipsoid Method optimization --- ### Robust Optimization Formulation - Consider $$\begin{array}{ll} \text{minimize} & \sup_{{\color{coral}q} \in \mathbb{Q} } {\color{purple}f_0}({\color{darkblack}x},{\color{coral}q}), \\ \text{subject to} & {\color{purple}f_j}({\color{darkblack}x},{\color{coral}q}) \leq 0 \qquad \\ & \forall {\color{coral}q} \in \mathbb{Q} \; \text{and} \; j = 1,2,\cdots,m, \\ \end{array}$$ where - ${\color{darkblack}x} \in \mathbb{R}^n$ represents a set of design variables (such as ${\color{darkblack}L}$, ${\color{darkblack}W}$), - ${\color{coral}q}$ represents a set of varying parameters (such as ${\color{coral}T_{OX}}$) - ${\color{purple}f_j} \leq 0$ represents the $j$th specification requirement (such as phase margin $\geq 60^\circ$). --- ### GP in Standard Form 📜 - We further assume that ${\color{purple}f_i}({\color{darkblack}x},{\color{coral}q})$'s are convex for all ${\color{coral}q} \in \mathbb{Q}$. - Geometric programming is an optimization problem that takes the following standard form: $$\begin{array}{lll} \text{minimize} & {\color{olive}p_0}({\color{darkblack}y}) & \\ \text{subject to} & {\color{olive}p_i}({\color{darkblack}y}) \leq 1, & i=1,\ldots,l \\ & {\color{firebrick}g_j}({\color{darkblack}y}) = 1, & j=1,\ldots,m \\ & {\color{darkblack}y}_k > 0,& k=1,\ldots,n , \end{array}$$ where - ${\color{olive}p_i}$'s are posynomial functions and ${\color{firebrick}g_j}$'s are monomial functions. --- ### Posynomial and Monomial Functions - A monomial function is simply: $${\color{firebrick}g}({\color{darkblack}y}_1,\ldots,{\color{darkblack}y}_n) = {\color{coral}c} {\color{darkblack}y}_1^{ {\color{coral}\alpha}_{1} } {\color{darkblack}y}_2^{ {\color{coral}\alpha}_{2} } \cdots {\color{darkblack}y}_n^{ {\color{coral}\alpha}_{n} }, \quad {\color{darkblack}y}_k > 0.$$ where - ${\color{coral}c}$ is non-negative and ${\color{coral}\alpha}_{k}\in {\mathbb{R} }$. - A posynomial function is a sum of monomial functions: $${\color{olive}p}({\color{darkblack}y}_1,\ldots,{\color{darkblack}y}_n) = \sum_{s=1}^{T}{{\color{coral}c_s} {\color{darkblack}y}_1^{ {\color{coral}\alpha}_{1,s} } {\color{darkblack}y}_2^{ {\color{coral}\alpha}_{2,s} } \cdots {\color{darkblack}y}_n^{ {\color{coral}\alpha}_{n,s} } }, \quad {\color{darkblack}y}_k > 0 ,$$ - A monomial can also be viewed as a special case of posynomial where there is only one term of the sum. --- ### GP in Convex Form - Many engineering problems can be formulated as a GP 🤔. - On Boyd's website there is a Matlab package "GGPLAB" and an excellent tutorial material 📚. - GP can be converted into a convex form by changing the variables ${\color{darkblack}x}_k = \log({\color{darkblack}y}_k)$ and replacing ${\color{olive}p_i}$ with $\log {\color{olive}p_i}$: $$\begin{array}{lll} \text{minimize} & \log {\color{olive}p_0}(\exp({\color{darkblack}x})) & \\ \text{subject to} & \log {\color{olive}p_i}(\exp({\color{darkblack}x})) \leq 0, & i=1,\ldots,l \\ & {\color{coral}a_j}^\mathsf{T} {\color{darkblack}x} + {\color{coral}b_j} = 0, & j=1,\ldots,m \end{array}$$ where - $\exp({\color{darkblack}x}) = (e^{ {\color{darkblack}x}_1}, e^{ {\color{darkblack}x}_2}, \cdots, e^{ {\color{darkblack}x}_n})$ - ${\color{coral}a}_j = ({\color{coral}\alpha}_{1,j}, \cdots, {\color{coral}\alpha}_{n,j})$ - ${\color{coral}b}_j = \log({\color{coral}c}_j)$ --- ### Robust GP 🛠️ .pull-left[ - GP in the convex form can be solved efficiently by interior-point methods 📈. - In robust version, coefficients ${\color{coral}c}_s$ are functions of ${\color{coral}q}$ 🔄. - The robust problem is still convex. Moreover, there is an infinite number of constraints 🔁. - Alternative approach: Ellipsoid Method 🫒. ] .pull-right[  The streetlights are bright, but are you sure this is where you dropped your key? 🔦🧐 ] --- count: false class: nord-light, middle, center ## Affine Arithmetic --- ### Example: Triangle Area 🔺 - Points ${\color{coral}p}$, ${\color{coral}q}$, and ${\color{coral}r}$ vary within 3 given rectangular regions - **Q**: What are the upper/lower bounds of $\triangle {\color{coral}p}{\color{coral}q}{\color{coral}r}$ area?  --- ### Method 1: Corner-based Approach - Calculate areas for all corner combinations - **Issues**: - Combinatorial explosion with many variables 🧩 - Worst-case may not occur at corners ⚠️ - Computationally expensive for high dimensions ⏳ --- ### Method 2: Monte Carlo Simulation 🎰 - Randomly sample points within variation ranges 🎲 - Calculate area for each sample - **Pros**: - More accurate with more samples - **Cons**: - Time-consuming ⏱️ - No guarantees of finding true bounds --- class: center, middle ## Interval vs. Affine Arithmetic --- ### 🧮 Method 3: Interval Arithmetic (IA) **Example Calculation**: - px = [2, 3], py = [3, 4] - qx = [-5, -4], qy = [-6, -5] - rx = [6, 7], ry = [-5, -4] **Area Calculation**: - ((qx - px)(ry - py) - (qy - py)(rx - px)) / 2 - IA result: [33..61] (Actual: [36.5..56.5]) **Limitation**: Cannot handle variable correlations ❌ --- ### Method 4: Affine Arithmetic (AA) ➕ **Advantages**: - Handles first-order correlations ✅ - More accurate than IA (e.g., [35..57] in example) 🎯 - Faster than Monte Carlo for small variations ⚡ **Implementation**: - Public library: `libaffa.a`/YALAA - Supports: +, -, *, /, sin(), cos(), pow(), etc. --- ### What is Affine Arithmetic? 📚 **Affine Form**: $${\color{darkblack}\hat{x}} = {\color{darkblack}x_0} + {\color{darkblack}x_1} {\color{coral}\epsilon}_1 + \ldots + {\color{darkblack}x_n} {\color{coral}\epsilon}_n$$ where: - ${\color{coral}\epsilon}_i \in [-1, 1]$ (noise symbols) - ${\color{darkblack}x}_0$: central value - ${\color{darkblack}x}_i$: partial deviations **Example Conversion**: IA [3..4] → $3.5 + 0.5 {\color{coral}\epsilon}_1$ --- ### Geometry of AA 📐 .pull-left70[ - Affine forms that share noise symbols are dependent: - ${\color{darkblack}\hat{x}} = {\color{darkblack}x_0} + {\color{darkblack}x_1} {\color{coral}\epsilon}_1 + \ldots + {\color{darkblack}x_n} {\color{coral}\epsilon}_n$ - ${\color{darkblack}\hat{y}} = {\color{darkblack}y_0} + {\color{darkblack}y_1} {\color{coral}\epsilon}_1 + \ldots + {\color{darkblack}y_m} {\color{coral}\epsilon}_m$ - The region containing (x, y) is: - ${\color{darkblack}Z} = \{({\color{darkblack}x}, {\color{darkblack}y}) : {\color{coral}\epsilon}_i \in [-1, 1]\}$ - This region is a centrally symmetric convex polygon called "zonotope". ] .pull-right30[  ] --- ### Enabling Technologies 💻 **C++ Template Metaprogramming**: ```cpp template <typename Tp> Tp area(const Tp& px, const Tp& qx, const Tp& rx, const Tp& py, const Tp& qy, const Tp& ry) { return ((qx-px)*(ry-py) - (qy-py)*(rx-px)) / 2; } ``` - Same code works for both `double` and `AAF` types - Enables code reuse with minimal modifications --- ### Open-Loop Gain Example 📈 - Open-loop gain ${\color{firebrick}A_v}$ can be approximated as a monomial function: $${\color{firebrick}A_v} = \frac{2 {\color{coral}C_{ox}} }{({\color{coral}\lambda_n} + {\color{coral}\lambda_p})^2} \sqrt{{\color{coral}\mu_n} {\color{coral}\mu_p} \frac{{\color{darkblack}W_1} {\color{darkblack}W_6}}{{\color{darkblack}L_1} {\color{darkblack}L_6} {\color{firebrick}I_1} {\color{firebrick}I_6} } }$$ where ${\color{firebrick}I_1}$ and ${\color{firebrick}I_6}$ are monomial functions. - Corresponding C++ code fragment: ```cpp // Open Loop Gain monomial<aaf> OLG = 2*COX/square(LAMBDAN+LAMBDAP)* sqrt(KP*KN*W[1]/L[1]*W[6]/L[6]/I1/I6); ``` --- ### Applications of AA 🛠️ - Analog Circuit Sizing 🎛️ - Worst-Case Timing Analysis ⏱️ - Statistical Static Timing Analysis - Parameter Variation Interconnect Modeling [CMU02] 🔌 - Clock Skew Analysis ⏰ - Bounded Skew Clock Tree Synthesis 🕰️ --- ### Limitations of AA ⚠️ **Functional Limitations**: - No iterative methods (no fixed point in AA) 🔄 - No multiplicative inverse (no LU decomposition) ➗ - No total ordering (can't sort) 🔢 **Mathematical Limitations**: - Only handles linear correlations - Poor approximation of `abs(x)` near zero ❌ Fortunately the ellipsoid method is one of the few algorithms that works with AA ✅ --- ### Affine Arithmetic How to find $\sup_{{\color{coral}q} \in {\mathbb{Q} } } {\color{purple}f_j}({\color{darkblack}x},{\color{coral}q})$ efficiently? - $\sup_{{\color{coral}q} \in {\mathbb{Q} } } {\color{purple}f_j}({\color{darkblack}x},{\color{coral}q})$ is in general difficult to obtain. - Provided that variations are small or nearly linear, we propose using Affine Arithmetic (AA) to solve this problem. - Features of AA: - Handle correlation of variations by sharing *noise symbols*. - Enabling technology: template and operator overloading features of C++. - A C++ package "YALAA" is publicly available. --- ### AA for Worst Case Analysis 🔍 - An uncertain quantity is represented in an affine form (AAF): $${\color{coral}\hat{a}} = {\color{darkblack}a_0} + {\color{darkblack}a_1} {\color{coral}\varepsilon}_1 + {\color{darkblack}a_2} {\color{coral}\varepsilon}_2 + \cdots + {\color{darkblack}a_k} {\color{coral}\varepsilon}_k = {\color{darkblack}a_0} + \sum_{i=1}^{k} {\color{darkblack}a_i} {\color{coral}\varepsilon}_i,$$ where - ${\color{coral}\varepsilon}_i \in [-1, 1]$ is called *noise symbol*. - Exact results for affine operations (${\color{coral}\hat{a}}+{\color{coral}\hat{b}}$, ${\color{coral}\hat{a}}-{\color{coral}\hat{b}}$ and ${\color{coral}\alpha} \cdot {\color{coral}\hat{a}}$) - Results of non-affine operations (such as ${\color{coral}\hat{a}} \cdot {\color{coral}\hat{b}}$, ${\color{coral}\hat{a}}/{\color{coral}\hat{b}}$, $\max({\color{coral}\hat{a}}, {\color{coral}\hat{b}}), \log({\color{coral}\hat{a}})$) are *approximated* in an affine form. - AA has been applied to a wide range of applications recently when process variations are considered. --- ### AA for Optimization ⚙️ In our robust GP problem: AA is used in worst-case scenario analysis to determine the maximum and minimum values of a function subject to variations in its inputs. First, represent every elements in ${\color{coral}q}$ in affine forms. For each ellipsoid iteration, ${\color{purple}f}({\color{darkblack}x}_c,{\color{coral}q})$ is obtained by *approximating* ${\color{purple}f}({\color{darkblack}x}_c,{\color{coral}\hat{q}})$ in an affine form: $${\color{purple}\hat{f}} = {\color{darkblack}f_0} + {\color{darkblack}f_1} {\color{coral}\varepsilon}_1 + {\color{darkblack}f_2} {\color{coral}\varepsilon}_2 + \cdots + {\color{darkblack}f_k} {\color{coral}\varepsilon}_k.$$ Then the maximum of $\hat{f}$ is determined by: $${\color{coral}\varepsilon}_j = \left\{ \begin{array}{ll} +1 & \qquad \text{if} \; {\color{purple}f}_j > 0 \\ -1 & \qquad \text{if} \; {\color{purple}f}_j < 0 \end{array} \right. \quad j=1, \cdots, k.$$ --- ### Robust Circuit Sizing Problem ⚖️ - Given a circuit topology and a set of specification requirements: .font-sm.mb-xs[ | Constraint | Spec. | Units | | -------------- | ---------- | ---------- | | Device Width | $\geq 2.0$ | $\mu$m | | Device Length | $\geq 1.0$ | $\mu$m | | Estimated Area | minimize | $\mu$m$^2$ | | $\vdots$ | $\vdots$ | $\vdots$ | | CMRR | $\geq 75$ | dB | | Neg. PSRR | $\geq 80$ | dB | | Power | $\leq 3$ | mW | ] - Find the worst-case design variable values that meet the specification requirements and optimize circuit performance 🎯. --- .pull-left70[  ] --- ### Performance Specifications 🎯 .column-2.font-sm.mb-xs[ | Constraint | Spec. | Units | |------------------|---------------|---------------| | Device Width | ≥ 2.0 | μm | | Device Length | ≥ 1.0 | μm | | Area | minimize | μm² | | Input CM Voltage | [0.45,0.55] | × VDD | | Output Range | [0.1,0.9] | × VDD | | Gain | ≥ 80 | dB | | Unity Gain Freq. | ≥ 50 | MHz | | Phase Margin | ≥ 60 | degrees | | Slew Rate | ≥ 50 | V/$\mu$s | | CMRR | ≥ 75 | dB | | Neg. PSRR | ≥ 80 | dB | | Power | $\leq 3$ | mW | | Noise, Flicker | $\leq 800$ | nV/Hz$^{0.5}$ | ] --- ### Design Variable Results .column-2.font-sm.mb-xs[ | Variable | Units | GGPLAB | Our | Robust | | ---------- | ------ | ------- | ----- | ------ | | ${\color{darkblack}W_1}$ | μm | 44.8 | 44.9 | 45.4 | | ${\color{darkblack}W_8}$ | μm | 3.94 | 3.98 | 3.8 | | ${\color{darkblack}W_{10}}$ | μm | 2.0 | 2.0 | 2.0 | | ${\color{darkblack}W_{13}}$ | μm | 2.0 | 2.0 | 2.1 | | ${\color{darkblack}L_1}$ | μm | 1.0 | 1.0 | 1.0 | | ${\color{darkblack}L_8}$ | μm | 1.0 | 1.0 | 1.0 | | ${\color{darkblack}L_{10}}$ | μm | 1.0 | 1.0 | 1.0 | | ${\color{darkblack}L_{13}}$ | μm | 1.0 | 1.0 | 1.0 | | ${\color{coral}A}$ | N/A | 10.4 | 10.3 | 12.0 | | ${\color{coral}B}$ | N/A | 61.9 | 61.3 | 69.1 | | ${\color{coral}C_c}$ | pF | 1.0 | 1.0 | 1.0 | | ${\color{coral}I_{bias}}$ | μA | 6.12 | 6.19 | 5.54 | ] --- ### Performance Comparison 🏆 .font-sm.mb-xs[ | Performance (units) | Spec. | Std. | Robust | | ------------------------------ | ------------ | ------------ | ------------ | | Estimated Area ($\mu$m$^2$) | minimize | 5678.4 | 6119.2 | | Output Range (x ${\color{coral}V_{DD}}$) | [0.1, 0.9] | [0.07, 0.92] | [0.07, 0.92] | | Comm Inp Range (x ${\color{coral}V_{DD}}$) | [0.45, 0.55] | [0.41, 0.59] | [0.39, 0.61] | | Gain (dB) | $\geq 80$ | 80 | [80.0, 81.1] | | Unity Gain Freq. (MHz) | $\geq 50$ | 50 | [50.0, 53.1] | | Phase Margin (degree) | $\geq 60$ | 86.5 | [86.1, 86.6] | | Slew Rate (V/$\mu$s) | $\geq 50$ | 64 | [66.7, 66.7] | | CMRR (dB) | $\geq 75$ | 77.5 | [77.5, 78.6] | | Neg. PSRR (dB) | $\geq 80$ | 83.5 | [83.5, 84.6] | | Power (mW) | $\leq 3$ | 1.5 | [1.5, 1.5] | | Noise, Flicker (nV/Hz$^{0.5}$) | $\leq 800$ | $600$ | [578, 616] | ] --- ### Conclusions 🎯 - Fast computation (<1 sec on 3GHz CPU) ⚡ - Guaranteed worst-case performance 🛡️ - Robust GP formulation handles variations 📈 - Ellipsoid method provides reliable solutions 🏗️ --- ### Final Comments 💭 - AA excels for small/near-linear variations ➕ - Combines well with zonotope geometry 🧩 - Not universal - incompatible with: - Iterative methods 🔄 - Matrix inversions ➗ - Best combined with other methods (Monte Carlo) 🔄 --- count: false class: nord-dark, middle, center .pull-left[ ## Q&A 🎤 ] .pull-right[  ]