Electronics Guide

Core Analysis Types

SPICE is a small set of numerical experiments performed on one netlist. Every higher-level study described later in this article, from Monte Carlo sampling to post-layout verification, is assembled from these primitives, so choosing the right one is the first decision in any simulation plan.

Static and Small-Signal Analyses

• Operating point (.OP): Solves the nonlinear circuit equations with every capacitor open and every inductor shorted, reporting node voltages, branch currents, and the linearized small-signal parameters of each device. Every other analysis begins from this solution, so an operating point that will not converge blocks the entire simulation
• DC sweep (.DC): Repeats the operating-point solution while stepping a source, a parameter, or the temperature. Transfer curves, input and output ranges, and current-mirror compliance all come from this analysis
• AC analysis (.AC): Linearizes the circuit about the operating point and solves the complex node equations at each frequency. Gain, phase, input and output impedance, and stability margins come from AC analysis, which by construction sees no slew-rate limit, no clipping, and no other nonlinearity
• Noise analysis (.NOISE): Sums the thermal, shot, and flicker noise contributed by every device at a chosen output and refers the total back to a designated input. It runs on the same linearization as AC analysis, so its results hold only where the circuit is genuinely small-signal about the operating point
• Transfer function and sensitivity (.TF, .SENS): Report small-signal gain together with input and output resistance, and the derivative of a chosen output with respect to each circuit and model parameter. Sensitivity results rank the parameters that a later Monte Carlo campaign most needs to exercise

Large-Signal and Time-Domain Analyses

• Transient analysis (.TRAN): Integrates the circuit equations forward in time, capturing slew limiting, clipping, switching, oscillation, and startup behavior that no linearized analysis can show. Its accuracy rests on the integration method, the timestep control, and the tolerances discussed under convergence below
• Fourier and distortion analyses (.FOUR, .DISTO): Extract harmonic content from a transient result, or compute distortion products directly from the small-signal derivatives of the device models. Distortion figures obtained either way are only as trustworthy as the higher-order continuity of the underlying compact model
• Parameter stepping (.STEP): Wraps any analysis above in an outer loop over a component value, a model parameter, a supply voltage, or a temperature, producing the families of curves on which design exploration depends
• Choosing the analysis: Prefer the cheapest analysis that answers the question. A phase margin is an AC result, an overload recovery time is a transient result, and a bias error is a DC result. Reaching for transient analysis where AC analysis would serve costs hours of runtime and buries the answer in a waveform

Device Model Fundamentals

At the heart of any SPICE simulation lies the device model, a mathematical representation of how a physical component behaves electrically. The accuracy of simulation results depends critically on the quality and appropriateness of these models.

Semiconductor Device Models

SPICE includes built-in models for fundamental semiconductor devices:

• Diode models: The basic diode model captures the exponential I-V relationship, junction capacitance variation with voltage, and reverse recovery characteristics. Key parameters include saturation current (IS), ideality factor (N), series resistance (RS), and junction capacitance parameters (CJO, VJ, M)
• Bipolar junction transistor (BJT) models: The SPICE Gummel-Poon model extends the basic Ebers-Moll equations to include high-injection effects, Early voltage, base resistance, and charge storage. Parameters such as BF (the forward current gain, β_F), IS (saturation current), VAF (forward Early voltage), and the forward and reverse transit times TF and TR define transistor behavior across operating conditions. Gummel-Poon remains the bipolar model built into nearly every simulator, but it does not represent self-heating, avalanche multiplication, the parasitic substrate transistor, or the graded profiles of silicon-germanium heterojunction devices. Processes that need those effects ship VBIC, Mextram, or HICUM instead, all maintained as public standards by the Compact Model Coalition. Treat Gummel-Poon as the general-purpose default rather than as the current state of bipolar modeling
• MOSFET models: Field-effect transistor modeling has evolved through several generations, selected in a netlist by a LEVEL keyword whose numbering is specific to each simulator. LEVEL 1 is the Shichman-Hodges square law, LEVEL 2 a more detailed analytical long-channel model, and LEVEL 3 a semi-empirical short-channel model. All three are teaching and hand-check models, and none describes a modern process; a page of LEVEL 3 parameters is no substitute for a foundry model. Production work uses the compact models the foundry delivers, which the same LEVEL keyword selects at higher numbers (HSPICE, for instance, reaches BSIM4 at LEVEL 54). The BSIM (Berkeley Short-channel IGFET Model) family captures short-channel effects, mobility degradation, and gate leakage: BSIM4 stands behind most bulk planar CMOS process design kits, BSIM-CMG covers multi-gate FinFET and gate-all-around devices, and BSIM-IMG addresses ultra-thin-body silicon-on-insulator transistors. Charge-based and surface-potential formulations, notably BSIM-BULK and PSP, describe weak, moderate, and strong inversion with a single continuous expression and are preferred where moderate-inversion bias points and distortion must be predicted accurately. The Compact Model Coalition maintains these as public standards so that one parameter set behaves identically in competing simulators
• JFET models: Junction field-effect transistor models capture the quadratic relationship between drain current and gate-source voltage, along with channel-length modulation and capacitance characteristics

Model Parameters and Extraction

Device model parameters are determined through a process called parameter extraction:

• Measurement-based extraction: Physical devices are characterized through careful electrical measurements. DC characteristics, capacitance-voltage (C-V) measurements, and transient response data provide the basis for determining model parameters
• Process-based parameters: Semiconductor foundries provide model parameters derived from their manufacturing process characterization. These foundry-provided models, often called PDKs (Process Design Kits), ensure simulation accuracy for the specific fabrication process
• Parameter optimization: Automated optimization routines adjust model parameters to minimize the difference between measured device behavior and model predictions across multiple operating conditions
• Model validation: Extracted models must be validated against independent measurements to ensure they accurately predict device behavior across the intended operating range

Passive Component Models

Accurate passive component models are equally important for reliable simulation:

• Resistor models: Beyond ideal resistance, practical resistor models include temperature coefficients (TC1, TC2), voltage coefficients for voltage-dependent resistance, and parasitic capacitance for high-frequency applications
• Capacitor models: Real capacitors exhibit equivalent series resistance (ESR), equivalent series inductance (ESL), voltage coefficient, and temperature dependence. These parasitics become critical in power supply decoupling and high-frequency filtering applications
• Inductor models: Inductor models incorporate series resistance (DC and frequency-dependent), interwinding capacitance, and core losses for magnetic components. The frequency-dependent impedance behavior is crucial for RF and power applications
• Transmission line models: For high-frequency circuits and PCB interconnects, transmission line models capture propagation delay, characteristic impedance, and loss characteristics

Subcircuit Development

Subcircuits allow complex components and circuit blocks to be encapsulated and reused, promoting hierarchical design and enabling accurate modeling of multi-terminal devices like operational amplifiers and voltage references.

Subcircuit Structure and Syntax

A SPICE subcircuit definition consists of:

• Subcircuit header: The .SUBCKT statement defines the subcircuit name and its external nodes. These nodes serve as the interface between the subcircuit and the external circuit
• Internal components: The body of the subcircuit contains all the components that make up the model, using a combination of primitive elements (resistors, capacitors, controlled sources) and potentially other subcircuits
• Parameters: Subcircuits can accept parameters that allow a single definition to represent a family of components with different values. Default parameter values ensure the subcircuit functions without explicit parameter specification
• Subcircuit termination: The .ENDS statement marks the end of the subcircuit definition

Operational Amplifier Subcircuits

Op-amp subcircuits must capture critical performance characteristics:

• Open-loop gain and bandwidth: The frequency-dependent gain is typically modeled using one or more poles, with the dominant pole determining the gain-bandwidth product
• Input characteristics: Input bias current, input offset voltage, input capacitance, and common-mode input range are modeled through appropriate combinations of current sources, voltage sources, and capacitors
• Output characteristics: Output resistance, output current limiting, and output voltage swing are captured to predict behavior when driving various loads
• Slew rate: The maximum rate of output voltage change is modeled using nonlinear elements that limit charging current to internal compensation capacitors
• Power supply effects: Power supply rejection ratio (PSRR) and supply current variation are included in more sophisticated models

Hierarchical Design Practices

Effective subcircuit development follows established practices:

• Level of detail: Choose the appropriate complexity level for the application. A simple macromodel may suffice for system-level simulation, while transistor-level detail is needed for detailed analog design
• Node naming conventions: Consistent, descriptive node names improve readability and facilitate debugging. Internal nodes should clearly indicate their function
• Documentation: Include comments describing the subcircuit purpose, limitations, and required parameters. Document the source of model parameters and validation status
• Library organization: Group related subcircuits into libraries with consistent naming conventions. Version control ensures traceability of model changes

Behavioral Modeling Techniques

Behavioral modeling uses mathematical expressions to describe component behavior rather than detailed physical structures. This approach enables efficient simulation of complex systems and modeling of components whose internal structure is unknown or irrelevant.

Voltage and Current Controlled Sources

SPICE provides linear controlled sources as building blocks:

• Voltage-controlled voltage source (VCVS): Output voltage is proportional to an input voltage. Used for modeling gain stages, buffers, and isolation amplifiers
• Voltage-controlled current source (VCCS): Output current is proportional to an input voltage. Essential for modeling transconductance amplifiers and operational transconductance amplifiers (OTAs)
• Current-controlled voltage source (CCVS): Output voltage is proportional to an input current. Used for transimpedance amplifiers and current-sensing applications
• Current-controlled current source (CCCS): Output current is proportional to an input current. Models current mirrors and current amplifiers

Nonlinear Behavioral Sources

Modern SPICE variants support arbitrary mathematical expressions:

• Arbitrary behavioral sources: The B-element of ngspice and LTspice allows a voltage or current to be specified as an arbitrary function of other circuit quantities, while PSpice and HSPICE express the same thing through a controlled source carrying an expression in place of a constant gain. This enables modeling of limiters, multipliers, dividers, and other nonlinear functions
• Mathematical functions: Standard mathematical functions (exp, log, sin, cos, abs, sqrt, etc.) and conditional expressions (if-then-else) enable sophisticated behavioral descriptions
• Table-based models: Lookup tables with interpolation model components characterized by measured data when analytical expressions are impractical
• Laplace transfer functions: Frequency-domain transfer functions defined using Laplace notation enable efficient modeling of filters and control systems without explicit component implementation

Mixed-Level Modeling

Combining behavioral and device-level models optimizes simulation efficiency:

• System partitioning: Use behavioral models for blocks outside the primary area of interest while retaining full transistor-level detail where accuracy is critical
• Interface modeling: Behavioral models must accurately represent the interface characteristics (input/output impedance, bandwidth) that affect connected circuitry
• Validation requirements: Behavioral models should be validated against transistor-level simulations or measurements to ensure they adequately represent the actual circuit behavior
• Simulation speed considerations: Behavioral models can dramatically reduce simulation time for large systems, enabling more extensive design exploration

Convergence Issues and Solutions

Convergence problems occur when SPICE cannot find a consistent solution to the circuit equations. Understanding the causes and solutions for convergence issues is essential for effective simulation.

DC Operating Point Convergence

Finding the DC operating point is often the most challenging convergence task:

• Newton-Raphson algorithm: SPICE uses iterative Newton-Raphson methods to solve nonlinear equations. The algorithm requires good initial guesses and may fail to converge for poorly conditioned problems
• Source stepping: Gradually ramping power supplies from zero to their final values can help establish initial conditions. Simulators apply this automatically when the direct Newton-Raphson solution fails; ngspice, for example, attempts GMIN stepping first and falls back to source stepping. HSPICE additionally offers the RAMPTIME option for an explicit supply ramp, and any simulator accepts explicitly ramped sources
• Node-set and initial conditions: The .NODESET statement provides initial voltage guesses to aid convergence. The .IC statement sets initial conditions for transient analysis starting points
• GMIN stepping: SPICE adds small conductances (GMIN) between nodes to prevent floating node problems. GMIN stepping gradually reduces these conductances to find the true solution

Common Causes of Convergence Failure

Several circuit characteristics commonly cause convergence problems:

• Positive feedback loops: Bistable circuits like latches and comparators with hysteresis have multiple valid operating points, confusing the DC solver. Breaking feedback loops during DC analysis or using initial conditions can resolve this
• Floating nodes: Nodes with no DC path to ground or a supply create infinite impedance that the solver cannot handle. Adding large resistances to ground provides a DC path without significantly affecting AC behavior
• Extreme impedance ratios: Very large differences in component values (such as picofarad capacitors with teraohm resistors) create numerical conditioning problems. Scaling components or using appropriate simulator tolerances helps
• Discontinuous models: Models with abrupt transitions or discontinuities can cause the Newton-Raphson algorithm to oscillate. Smoother model transitions or tighter convergence tolerances may be required

Transient Analysis Convergence

Time-domain simulations present additional challenges:

• Time step control: Adaptive time stepping adjusts step size based on signal activity. Rapid transitions require small steps, while steady-state periods allow larger steps. The TRTOL parameter, which defaults to 7, scales the local truncation error estimate that decides whether a step is accepted; it represents the factor by which SPICE is believed to overstate the true truncation error, and circuits containing ideal switches usually need it reduced
• Breakpoints: Piecewise-linear sources and switching events create breakpoints where the simulator must evaluate the solution. Excessive breakpoints slow simulation and may cause convergence problems
• Integration method: SPICE offers multiple numerical integration methods (trapezoidal, Gear) with different stability and accuracy characteristics. Gear methods are more stable for stiff systems but may exhibit numerical damping
• Maximum time step: Setting TMAX limits the maximum time step, ensuring adequate resolution of high-frequency signals that might otherwise be aliased

Convergence Parameter Adjustment

Several simulator options govern convergence, and their interaction rewards understanding rather than guesswork. Newton-Raphson iteration stops when successive iterates agree to within a relative tolerance plus an absolute floor: a node voltage has converged when the change between iterations is smaller than RELTOL times its magnitude or smaller than VNTOL, whichever is larger, and a branch current when the change is smaller than RELTOL times its magnitude or smaller than ABSTOL. The absolute floors exist so that a quantity passing through zero does not demand infinite relative precision. The defaults below are those of Berkeley SPICE3 and the simulators that inherit from it, such as ngspice; commercial tools ship comparable numbers.

• ABSTOL: The absolute current floor, 1 picoampere by default. It is the option most often wrong for precision analog work: a photodiode front end or a gate-leakage measurement whose signal currents are femtoamperes will see the solver declare numerical noise converged unless ABSTOL is tightened to match
• VNTOL: The absolute voltage floor, 1 microvolt by default. Circuits resolving nanovolt-scale offsets need it reduced in the same way
• RELTOL: The relative tolerance, 0.001 by default, or one part in a thousand. Loosening it toward 0.01 often rescues a stubborn transient run, but it loosens every converged quantity in the circuit and degrades accuracy directly, so treat it as a diagnostic step rather than a fix. Tightening it below roughly 0.0001 rarely pays and can make convergence impossible
• ITL1, ITL2, and ITL4: Iteration limits of 100 for the DC operating point, 50 for the DC transfer curve, and 10 per timestep in transient analysis. Exceeding ITL4 is not an error: the simulator cuts the timestep and retries, so raising ITL4 to several tens can carry a run past a difficult switching edge instead of stalling it in ever-smaller steps
• GMIN: The minimum conductance placed across every nonlinear device junction, 1 picosiemens by default, equivalent to a 1 teraohm leakage path. Larger values improve convergence but impose a real leakage error, which matters in sample-and-hold circuits, integrators, and other very high impedance nodes
• CHGTOL: The charge tolerance, 10⁻¹⁴ coulombs by default, which enters the transient timestep control alongside TRTOL. Circuits that switch around capacitors usually want both values reduced together

Change these options deliberately and record the change. A result obtained with loosened tolerances is a different result, and a simulation that converges only after RELTOL is relaxed usually reports a modeling problem rather than a numerical one.

Monte Carlo Analysis

Monte Carlo analysis assesses circuit performance in the presence of random component variations, providing statistical insight into manufacturing yield and performance distributions.

Monte Carlo Methodology

The fundamental approach involves repeated simulations with varied parameters:

• Parameter variation: Each simulation run samples component parameters from specified probability distributions. Gaussian (normal) and uniform distributions are most commonly used
• Run count selection: The number of runs determines statistical confidence. Hundreds or thousands of runs may be needed for accurate tail distribution estimation. The required count increases with the confidence level and distribution complexity
• Performance metric extraction: For each run, key performance parameters (gain, bandwidth, offset, power consumption) are measured and recorded
• Statistical analysis: The collection of results is analyzed to determine mean, standard deviation, and distribution shape. Histograms visualize the performance distribution

Variation Types

Different variation mechanisms require distinct modeling approaches:

• Global variations: Process variations that affect all devices on a die similarly. Examples include oxide thickness variations across a wafer. Global parameters are varied once per Monte Carlo run and applied to all instances
• Local (mismatch) variations: Random variations between nearby identical devices on the same die. These cause offset voltages in differential pairs and current mismatch in current mirrors. Each device instance receives independent random variations
• Lot-to-lot and wafer-to-wafer variations: Variations between manufacturing lots, and between wafers within a lot. Statistical foundry models normally fold these into the same global distribution that is sampled once per Monte Carlo run, so the global term already covers them rather than their forming a separate mechanism. The named process corners are simply the tails of that same distribution given convenient labels
• Environmental variations: Temperature and supply voltage variations may be included in Monte Carlo analysis to assess combined effects

Interpreting Monte Carlo Results

Proper interpretation enables design decisions:

• Yield estimation: The fraction of runs meeting all specifications indicates expected manufacturing yield. Be precise about what a sigma target means, because two conventions circulate under the same name. In circuit design, six sigma means the specification limit sits six standard deviations from the mean, which for a two-sided Gaussian leaves roughly two failures per billion parts. The familiar figure of 3.4 defects per million belongs to the manufacturing convention, which permits the process mean to drift by one and a half sigma and therefore corresponds to a limit four and a half sigma out. Confusing the two understates the required margin by more than three orders of magnitude. Note also that a few thousand plain Monte Carlo runs cannot measure a six-sigma tail at all; targets that far out require the importance-sampling or worst-case-distance methods described below
• Sensitivity identification: Correlating performance variations with parameter variations identifies the dominant contributors to performance spread, guiding design optimization
• Design margin: The difference between mean performance and specification limits indicates design margin. Insufficient margin leads to yield loss
• Distribution shape: Non-Gaussian distributions may indicate nonlinear sensitivity to parameter variations or specification violations that depend on multiple parameters varying unfavorably

Efficient Monte Carlo Techniques

Advanced methods improve Monte Carlo efficiency:

• Latin hypercube sampling: This stratified sampling technique provides better coverage of the parameter space with fewer runs than simple random sampling
• Importance sampling: By weighting samples toward regions of parameter space that produce specification violations, importance sampling estimates tail probabilities with fewer runs
• Response surface methods: A limited number of SPICE simulations characterize the relationship between parameters and performance. Extensive statistical analysis is then performed on the simpler response surface model
• Worst-case distance analysis: Mathematical optimization finds the parameter combination that most nearly violates specifications, identifying the weakest point in the design

Process Corner Analysis

Process corner analysis evaluates circuit performance at extreme combinations of process parameters, representing the boundaries of expected manufacturing variation. This deterministic approach complements statistical Monte Carlo analysis.

Corners are pessimistic and non-conservative at the same time, and both halves of that statement matter. They are pessimistic because a named corner drives every global parameter to a distribution edge simultaneously, a joint event far less likely than any one parameter reaching its own limit, so a design that fails only at the extreme corners may still ship at acceptable yield. They are non-conservative because a corner model shifts every device on the die together and therefore leaves nominally matched devices identical to one another: a corner run produces exactly zero input offset voltage in a differential pair and exactly zero error in a current mirror. A circuit whose real failure mode is local mismatch can pass every corner in the matrix and still fail in volume. Corners bound global variation; only Monte Carlo bounds the local variation described under mismatch analysis below.

Traditional Process Corners

Standard corners characterize device speed extremes:

• Typical-Typical (TT): All devices operate at nominal process parameters. This represents the center of the process distribution and is used for nominal performance characterization
• Fast-Fast (FF): Both NMOS and PMOS transistors are at their fast (high current) extreme. This corner typically produces highest power consumption and fastest switching speeds
• Slow-Slow (SS): Both transistor types are at their slow extreme. This corner shows minimum speed and usually minimum power consumption
• Fast-Slow (FS) and Slow-Fast (SF): These skewed corners have one transistor type fast while the other is slow. They stress circuits sensitive to NMOS/PMOS ratio, such as inverter switching thresholds and push-pull output stages

Extended Corner Analysis

Additional corners address specific concerns:

• Temperature corners: Simulation at the temperature extremes of the target qualification grade reveals temperature-dependent behavior. Industrial parts are usually qualified from −40 to 85 degrees Celsius, while the automotive grades of AEC-Q100 run from −40 to 105 degrees Celsius for grade 2, −40 to 125 degrees for grade 1, and −40 to 150 degrees for grade 0. Junction temperature, not ambient, is what the simulation must reach, so self-heating has to be added to the specified ambient. Combining temperature with process corners creates a complete operating envelope
• Supply voltage corners: Simulating at minimum and maximum supply voltages shows voltage sensitivity. Combined with process and temperature, this creates the complete PVT (Process, Voltage, Temperature) corner matrix
• Passive component corners: Resistors, capacitors, and inductors have their own process variations. Combining active and passive corners ensures comprehensive coverage
• Reliability corners: Worst-case aging or stress conditions may be included as additional corners for lifetime performance verification

Corner Selection Strategies

Efficient corner analysis requires thoughtful selection:

• Performance-specific corners: Different performance metrics may be worst at different corners. Maximum delay typically occurs at slow/low-voltage/high-temperature, while minimum delay occurs at fast/high-voltage/low-temperature
• Reduced corner sets: A carefully selected subset of corners can provide good coverage with fewer simulations. The specific reduced set depends on the circuit topology and performance requirements
• Interpolation and extrapolation: Results at intermediate conditions can sometimes be estimated from corner simulations, though this must be validated for accuracy
• Monte Carlo at corners: Running Monte Carlo analysis at each process corner (rather than just typical) provides more complete statistical characterization but dramatically increases simulation count

Foundry Corner Models

Semiconductor foundries provide corner models specific to their process:

• Model file organization: Corner models are typically provided as separate model files or as parameters selecting the corner within a single file. Design kits configure the simulation environment for corner selection
• Statistical correlation: Foundry-provided corners are based on statistical analysis of actual manufacturing data, ensuring they represent realistic parameter combinations
• Corner definitions: The exact definition of corners (3-sigma, 4-sigma, or other) varies between foundries and should be understood when interpreting results
• Model updates: As process monitoring accumulates data, foundry models are updated. Design teams must manage model versions and understand the implications of model updates

Mismatch Analysis

Device mismatch, the random variation between nominally identical devices on the same die, critically affects analog circuit performance. Differential pairs, current mirrors, and precision analog circuits are particularly sensitive to mismatch.

Mismatch Mechanisms

Multiple physical mechanisms contribute to device mismatch:

• Random dopant fluctuation: The discrete nature of dopant atoms creates statistical variation in threshold voltage. A small modern transistor contains only a countable number of channel dopant atoms, so the fractional variation in that number is large. Random dopant fluctuation dominated threshold mismatch throughout the polysilicon-gate era and scales with device area according to the Pelgrom model
• Metal gate work-function variation: In high-k metal-gate processes the gate electrode is polycrystalline, and grains of different crystal orientation present different work functions to the channel beneath them. The resulting threshold variation is comparable to random dopant fluctuation at advanced nodes, and unlike dopant fluctuation it does not improve by lowering the channel doping
• Line edge roughness: Lithographic variations create roughness in polysilicon gate edges, causing effective gate length variations between devices
• Oxide thickness variation: Local variations in gate oxide thickness affect threshold voltage and transconductance
• Well proximity effects: Distance from well edges and other structures creates systematic variations that appear as mismatch when comparing devices at different locations

Mismatch Model Parameters

Foundry models characterize mismatch statistically. Threshold voltage follows the Pelgrom area law

σ(V_T) = A_VT/√(WL)

where A_VT is a process-dependent constant expressed in millivolt-micrometers and WL is the drawn gate area. Confirm which convention a foundry uses before sizing anything, because Pelgrom's original formulation writes the same law for the difference between two identically drawn neighbors rather than for one device, and the two readings differ by the square root of two. A_VT tracks gate-oxide thickness closely, so it has improved steadily with scaling, from tens of millivolt-micrometers on early thick-oxide generations to a few millivolt-micrometers on deep-submicron CMOS. The improvement is slower than the shrink that drives it: one published comparison found feature size falling by a factor of four between the 180 and 45 nanometer nodes while the matching coefficient improved by only about a factor of two, which is why analog devices cannot simply be scaled alongside digital ones.

Rearranging the law makes its cost concrete. The required area is the square of the ratio of A_VT to the target sigma, so on the single-device reading a process whose A_VT is 5 millivolt-micrometers needs 25 µm² of gate area, a square roughly 5 µm on a side, to hold one device's threshold sigma to 1 mV. A differential pair built from two such devices still shows an input offset sigma near 1.4 mV, larger by the square root of two. On an older process whose A_VT is 10 millivolt-micrometers the same 1 mV target costs 100 µm², an area a precision op-amp input stage can barely afford. Each further halving of the offset costs four times the area, four times the gate capacitance, and the bandwidth that goes with it. That quadratic exchange rate governs the economics of precision analog design, and it is why chopping, auto-zeroing, and digital calibration exist.

The current factor obeys the same area law

σ(β)/β = A_β/√(WL)

Here β is the MOSFET current factor, the product of carrier mobility, gate-oxide capacitance per unit area, and the width-to-length ratio. It is not the bipolar forward current gain that the Gummel-Poon parameter BF sets, although the mismatch literature and the bipolar literature both claim the letter β. This article follows each convention in its own place, so read the symbol in context.

Threshold and current-factor mismatch combine in the drain-current error of a matched pair. Where the two sources are independent, the fractional current error is

σ(ΔI_D/I_D)² = (g_m/I_D)²σ(ΔV_T)² + σ(Δβ/β)²

A complete foundry model adds a cross term of the form 2ρ(g_m/I_D)σ(ΔV_T)σ(Δβ/β), where ρ is the measured correlation coefficient between the two mismatch sources. A nonzero ρ moves the predicted current spread in whichever direction the correlation runs, so the coefficient belongs in the model file rather than in a designer's assumption. The expression also shows which term is worth attacking. Because g_m/I_D is largest in weak inversion and falls as roughly two divided by the overdrive voltage in strong inversion, threshold mismatch dominates a current mirror biased at low overdrive, while current-factor mismatch sets a floor that no further increase in overdrive can cross. Mirror overdrive is therefore a matching decision as much as a headroom decision.

Two further effects complete the picture:

• Distance-dependent mismatch: Pelgrom's full expression adds a term proportional to the square of the separation between the two devices, representing wafer-scale gradients in oxide thickness, doping, and stress. The area term is what device sizing buys; the distance term is what close placement and common-centroid layout buy, and no amount of area suppresses it
• Passive device mismatch: Resistor and capacitor matching improves with device area in the same way, and matching coefficients are specified separately for each resistor and capacitor type the process offers. Metal-insulator-metal and polysilicon-polysilicon capacitors routinely match far better than transistors do, which is why switched-capacitor circuits carry a precision that a transistor-based equivalent cannot reach

Simulation Approaches for Mismatch

Several simulation techniques evaluate mismatch effects:

• Monte Carlo mismatch simulation: Each device instance receives independent random parameter variations. Multiple runs build statistical distributions of mismatch-sensitive parameters like offset voltage
• Sensitivity analysis: Analytical or small-signal methods determine how circuit performance depends on individual device parameter variations, identifying critical matching requirements
• Worst-case mismatch: Intentionally setting matched devices to opposite extremes shows worst-case mismatch effects. This deterministic approach complements statistical analysis
• Layout-dependent mismatch: Some simulators incorporate layout information to calculate systematic mismatch from physical placement and proximity effects

Design for Matching

Circuit and layout techniques minimize mismatch impact:

• Device sizing: Larger devices exhibit better matching due to the area-dependent nature of random mismatch. The trade-off between matching and other constraints (speed, power, area) drives device sizing decisions
• Common-centroid layout: Interleaving matched devices in a common-centroid pattern cancels systematic gradients while randomizing spatial mismatch contributions
• Dummy devices: Placing inactive dummy structures adjacent to matched devices ensures identical processing environment, particularly for edge effects
• Orientation matching: Using consistent device orientation avoids systematic differences from directional process variations

Temperature Sweeps

Temperature profoundly affects semiconductor device behavior. Temperature sweep simulations characterize circuit performance across the operating temperature range, ensuring reliable operation from cold start to high-temperature stress conditions.

Temperature Effects on Devices

Key device parameters vary with temperature:

• Threshold voltage: MOSFET threshold voltage falls by roughly 1 to 2 millivolts for every degree Celsius of temperature rise. This moves bias points and logic thresholds across the operating range, and it is the mechanism behind the temperature inversion described below
• Carrier mobility: Mobility decreases with temperature as phonon scattering increases. This reduces transistor transconductance and current drive capability at elevated temperatures
• Current gain: The forward current gain of a bipolar transistor, the β_F that the Gummel-Poon parameter BF sets, rises with temperature. Bias networks that depend on base current drift accordingly, which is why emitter degeneration and current-mirror biasing are preferred where the operating point must hold across temperature
• Leakage currents: Subthreshold leakage and junction leakage increase exponentially with temperature, roughly doubling for every 8 to 10 degrees Celsius of rise. Leakage becomes the dominant power component at elevated temperatures
• Resistance: Metal resistance increases with temperature (positive temperature coefficient), while semiconductor resistor behavior depends on doping level and can have positive or negative temperature coefficient

Temperature Sweep Implementation

SPICE provides mechanisms for temperature analysis:

• .TEMP statement: Specifies one or more temperatures at which to run the simulation. Each temperature produces a separate set of results
• Stepped analysis: The .STEP command sweeps temperature continuously, showing performance variation as a function of temperature
• Temperature-dependent parameters: Model parameters include temperature coefficients that modify values based on simulation temperature. The reference temperature (TNOM) establishes the baseline
• Local temperature variation: Some simulators support different temperatures for different parts of the circuit, modeling thermal gradients within an IC

Critical Temperature Analysis Points

Specific analyses target temperature-sensitive behavior:

• Bias stability: Current mirrors, voltage references, and bias networks must maintain stable operating points across temperature. Temperature sweep shows drift and identifies compensation needs
• Speed variation: Two temperature effects pull in opposite directions. Falling carrier mobility slows a device as it heats, while a falling threshold voltage speeds it up. At the supply voltages of older processes the mobility term wins and the slow case is hot. At the reduced supplies of modern nodes the threshold term can win instead, and the slowest case moves to the cold extreme. This reversal is known as temperature inversion, and it is the reason both extremes must be simulated rather than assumed
• Power consumption: Leakage rises exponentially with temperature and can dominate the total at the hot extreme, while at the cold extreme it nearly vanishes and switching activity accounts for almost all consumption. The worst case for total power is therefore hot, but the worst case for the ratio of dynamic to static power is cold, and thermal design must check both
• Startup behavior: Cold-start conditions may stress circuits differently than steady-state operation. Transient simulation at temperature extremes verifies proper startup

Temperature Compensation Design

Understanding temperature effects enables compensation:

• Bandgap references: Combining positive and negative temperature coefficient voltages creates temperature-independent references. Simulation across temperature verifies compensation effectiveness
• PTAT and CTAT currents: Proportional-to-absolute-temperature (PTAT) and complementary-to-absolute-temperature (CTAT) currents enable temperature-compensated biasing
• Thermal shutdown: Protection circuits must activate before damage temperatures are reached. Temperature sweep simulation verifies threshold accuracy and protection circuit response
• Temperature sensors: On-chip temperature sensing enables dynamic compensation. Sensor accuracy must be verified across the measurement range

Post-Layout Simulation with Parasitics

Physical layout introduces parasitic resistances, capacitances, and inductances not present in schematic-level simulation. Post-layout simulation includes these extracted parasitics to verify that the physical implementation meets performance requirements.

Parasitic Extraction

Specialized tools extract parasitics from layout geometry:

• Resistance extraction: Metal and polysilicon routing contribute series resistance. Via and contact resistance add to the total path resistance. Sheet resistance values from the process characterize each layer
• Capacitance extraction: Parallel plate capacitance between overlapping metal layers, fringe capacitance at conductor edges, and coupling capacitance between adjacent conductors are all extracted. Accuracy requirements determine the extraction methodology sophistication
• Inductance extraction: For high-frequency and high-speed circuits, mutual and self-inductance of routing becomes significant. Full electromagnetic extraction may be required for accurate modeling
• Extraction accuracy vs. speed: Simplified extraction methods run quickly but may miss important parasitics. Field-solver-based extraction provides higher accuracy at computational cost

Parasitic Netlist Formats

Extracted parasitics are represented in various formats:

• SPEF (Standard Parasitic Exchange Format): An industry-standard format that represents parasitics as lumped RC networks. Widely supported by extraction and simulation tools
• DSPF (Detailed Standard Parasitic Format): A SPICE-compatible format that describes the full distributed RC network of each net, including instance pin and node detail. DSPF files are typically far larger than the equivalent SPEF, which maps names to integers to conserve space
• Expanded SPICE netlist: Parasitics can be back-annotated directly into the SPICE netlist as explicit R and C elements. This creates larger netlists but ensures compatibility with any SPICE simulator
• Reduced-order models: For very large nets, model reduction techniques create compact representations that capture essential behavior with fewer elements

Post-Layout Analysis Types

Different analyses reveal various parasitic effects:

• RC delay analysis: Distributed RC networks in routing paths add delay to signal propagation. Critical paths must be checked for timing margin degradation
• IR drop analysis: Resistive power distribution networks cause voltage drop under load. Power grid integrity must be verified to ensure adequate voltage at all circuit locations
• Crosstalk analysis: Coupling capacitance between signal lines creates crosstalk. Victim nets may experience glitches or delay variation from aggressor activity
• Substrate coupling: Signals can couple through the conductive substrate, affecting sensitive analog circuits. Substrate extraction and analysis identifies potential coupling paths

Correlation with Silicon

Post-layout simulation must correlate with measured results:

• Extraction calibration: Extraction tools are calibrated against test structures measured on silicon. Accurate layer thickness and material properties are essential
• Model-to-hardware correlation: Comparing simulated and measured results identifies modeling gaps. Systematic differences indicate extraction or model errors requiring correction
• Guard-banding: Uncertainty in parasitic extraction may require design margin beyond what simulation predicts. Experience with a specific process guides appropriate margins
• Continuous improvement: Correlation data feeds back to improve extraction accuracy and model parameters for future designs

Advanced Simulation Techniques

Beyond basic SPICE analysis, advanced techniques address specialized requirements for complex analog design.

Periodic Steady-State Analysis

Circuits with periodic excitation benefit from specialized analysis:

• PSS fundamentals: Periodic steady-state analysis directly finds the periodic operating point, avoiding long transient settling times. This is essential for oscillators, switched-capacitor circuits, and RF mixers
• Shooting method: The circuit is simulated for one period, and boundary conditions are adjusted until the final state matches the initial state
• Harmonic balance: Frequency-domain techniques solve for the Fourier coefficients of all circuit waveforms, efficiently handling circuits with strong nonlinearity
• Small-signal analysis on PSS: Once periodic steady-state is found, small-signal analyses (PAC, PXF, PNOISE) characterize behavior around the periodic operating point

Noise Analysis

Understanding circuit noise is critical for sensitive analog design:

• Small-signal noise analysis: SPICE calculates output noise spectral density from device noise sources. Input-referred noise characterizes amplifier noise performance
• Noise figure: For RF circuits, noise figure quantifies the degradation in signal-to-noise ratio through the circuit
• Periodic noise analysis: For switched or clocked circuits, periodic noise analysis accounts for the time-varying nature of noise transfer
• Flicker noise: 1/f noise, particularly important at low frequencies, requires accurate model parameters for correct prediction

Reliability Simulation

Long-term reliability effects can be simulated:

• Hot carrier injection: High-field stress near the drain degrades transistor parameters over time. Aging models predict parameter shifts based on operating conditions
• NBTI (Negative Bias Temperature Instability): PMOS threshold voltage shift under negative gate bias stress affects circuit performance over product lifetime
• Electromigration: Current density in metal lines causes material migration over time. Layout analysis identifies wires at risk of electromigration failure
• Accelerated aging simulation: Applying stress conditions and aging models predicts circuit performance at end-of-life, ensuring adequate design margin

Multi-Domain Simulation

Complex systems require simulation across multiple physical domains:

• Electro-thermal simulation: Self-heating affects device parameters, which in turn affects power dissipation. Coupled electrical-thermal simulation captures these feedback effects
• Electromagnetic-circuit co-simulation: RF and high-speed circuits require combined electromagnetic field simulation and circuit simulation for accurate results
• Mixed-signal simulation: Analog and digital portions of mixed-signal ICs are simulated together, with appropriate representations for each domain
• System-level integration: SPICE-level blocks integrate into system simulators for verification of complete systems including control algorithms and digital processing

Best Practices and Methodology

Effective SPICE simulation requires disciplined methodology and attention to detail.

Simulation Planning

Thoughtful simulation planning improves efficiency:

• Define specifications: Clear performance specifications drive simulation requirements. Each specification maps to specific analyses and pass/fail criteria
• Create a simulation plan: Document the complete set of simulations required for design verification. Include stimulus conditions, analysis types, measurements, and acceptance criteria
• Prioritize analyses: Run critical simulations early to identify fundamental problems before investing time in detailed optimization
• Balance accuracy and speed: Use simplified models for initial exploration, progressing to full-accuracy simulations as the design matures

Testbench Development

Well-designed testbenches enable reliable simulation:

• Realistic stimulus: Apply stimulus waveforms that represent actual operating conditions, including realistic source impedances and loading
• Measurement circuits: Include ideal measurement circuits (peak detectors, RMS meters, frequency counters) that extract performance metrics without affecting circuit behavior
• Boundary conditions: Define appropriate power supply, temperature, and corner settings for each simulation
• Testbench reuse: Parameterize testbenches for reuse across different analyses and designs, reducing development time and ensuring consistency

Results Verification

Critical evaluation of simulation results prevents errors:

• Sanity checking: Verify that results are physically reasonable before accepting them. Unexpected values may indicate simulation errors
• Convergence verification: Check for warning messages indicating convergence difficulties that may affect accuracy
• Sensitivity analysis: Understanding how results depend on key parameters builds confidence in simulation validity
• Cross-checking: Where possible, verify results using hand calculations or alternative simulation approaches

Documentation and Archiving

Proper documentation ensures simulation value is preserved:

• Version control: Track all design files, models, and simulation scripts in version control systems
• Simulation logs: Maintain records of all simulations run, including setup parameters and results
• Design review materials: Prepare clear presentations of simulation results for design reviews
• Post-silicon correlation: Document comparison between simulation predictions and measured silicon performance to improve future modeling accuracy

Ultimately, SPICE is a tool for building design confidence, not a replacement for measurement. The strongest design flows treat simulation and silicon as complementary: models and testbenches are refined against measured data, and design margins absorb the residual uncertainty that no model fully captures. Verified models, planned analyses, and documented correlation are what turn a SPICE deck from a qualitative sketch into a quantitative predictor of circuit behavior.

Related Topics