Analog design preserves useful information carried by continuous voltage, current, charge, frequency, phase, or time. Even a “digital” chip depends on analog behavior at its boundaries and foundations: clock generation, power regulation, sensing, data conversion, wireline receivers, temperature monitors, and memory interfaces. The designer builds circuits whose gain, bandwidth, noise, linearity, stability, and power remain adequate despite device variation, parasitics, temperature, supply movement, and aging.

The central discipline is managing tradeoffs with physical models. More bias current can increase speed and reduce some noise, but it raises power and may reduce voltage headroom. Larger devices improve matching and flicker noise, but add capacitance and area. Higher loop gain improves accuracy until extra poles threaten stability. Analog design therefore proceeds as a sequence of budgets, calculations, simulations, layout decisions, and measured correlation—not as schematic invention followed by a cosmetic layout step.

<svg viewBox="0 0 960 390" xmlns="http://www.w3.org/2000/svg" font-family="-apple-system,Segoe UI,Roboto,sans-serif">
  <rect width="960" height="390" rx="18" fill="#0b1220"/>
  <text x="480" y="38" fill="#e5edf8" font-size="25" font-weight="700" text-anchor="middle">An analog signal chain is a sequence of budgets</text>
  <text x="480" y="64" fill="#93a4b8" font-size="14" text-anchor="middle">Signal quality is preserved only when every stage controls noise, range, and bandwidth</text>
  <g font-size="16" font-weight="700" text-anchor="middle">
    <rect x="42" y="112" width="140" height="78" rx="12" fill="#132238" stroke="#60a5fa" stroke-width="2"/><text x="112" y="145" fill="#bfdbfe">Sensor / input</text><text x="112" y="170" fill="#93a4b8" font-size="12" font-weight="400">µV to volts</text>
    <rect x="231" y="112" width="140" height="78" rx="12" fill="#251723" stroke="#f472b6" stroke-width="2"/><text x="301" y="145" fill="#fbcfe8">Amplify</text><text x="301" y="170" fill="#93a4b8" font-size="12" font-weight="400">gain + common mode</text>
    <rect x="420" y="112" width="140" height="78" rx="12" fill="#10251f" stroke="#34d399" stroke-width="2"/><text x="490" y="145" fill="#a7f3d0">Filter</text><text x="490" y="170" fill="#93a4b8" font-size="12" font-weight="400">band limit</text>
    <rect x="609" y="112" width="140" height="78" rx="12" fill="#292013" stroke="#fbbf24" stroke-width="2"/><text x="679" y="145" fill="#fde68a">Sample / ADC</text><text x="679" y="170" fill="#93a4b8" font-size="12" font-weight="400">quantize</text>
    <rect x="798" y="112" width="120" height="78" rx="12" fill="#1c1b2f" stroke="#a78bfa" stroke-width="2"/><text x="858" y="145" fill="#ddd6fe">Digital</text><text x="858" y="170" fill="#93a4b8" font-size="12" font-weight="400">estimate</text>
  </g>
  <g stroke="#64748b" stroke-width="3" fill="none"><path d="M182 151h41"/><path d="M371 151h41"/><path d="M560 151h41"/><path d="M749 151h41"/></g>
  <g fill="#64748b"><path d="M223 151l-9-7v14z"/><path d="M412 151l-9-7v14z"/><path d="M601 151l-9-7v14z"/><path d="M790 151l-9-7v14z"/></g>
  <rect x="110" y="250" width="740" height="90" rx="14" fill="#101927" stroke="#334155"/>
  <text x="140" y="281" fill="#67e8f9" font-size="16" font-weight="700">Shared constraints:</text><text x="300" y="281" fill="#cbd5e1" font-size="15">reference • clock • supply • temperature • process • layout</text>
  <text x="480" y="314" fill="#93a4b8" font-size="14" text-anchor="middle">An error added early is amplified downstream; an overloaded stage cannot be repaired later.</text>
</svg>

Specifications begin with the signal and its environment. Input range, source impedance, desired information bandwidth, interference, load, supply, temperature, and allowable latency define the problem. System architects then allocate gain, noise, linearity, and headroom across stages. If the first amplifier adds excessive noise, later filtering cannot recover the lost signal-to-noise ratio. If an early node clips, extra ADC resolution only describes the clipped waveform more precisely.

Dynamic range compares the largest acceptable signal with the smallest detectable one. In voltage form, (DR=20\log_{10}(V_{max}/V_{min})). The limits can come from noise, offset uncertainty, distortion, supply rails, device breakdown, or ADC range. Stating bandwidth and measurement conditions with dynamic range is essential; integrating noise over ten times more bandwidth raises total noise even when spectral density is unchanged.

MOS transistors translate bias current and geometry into transconductance, resistance, capacitance, and noise. Around an operating point, the small-signal drain-current change is approximately (i_d=g_m v_{gs}+g_{ds}v_{ds}). Intrinsic voltage gain is related to (g_m r_o). Modern short-channel devices offer high speed but reduced output resistance and limited voltage headroom, making cascoding and gain stacking harder.

The (g_m/I_D) method organizes inversion level and current efficiency. High (g_m/I_D) in weak or moderate inversion provides transconductance efficiently and supports low-voltage, low-power design, while stronger inversion can offer higher speed per device capacitance and different linearity. Characterized lookup tables from the target process are more reliable than long-channel hand equations for final sizing. Hand analysis remains valuable because it reveals sensitivities and provides checks on simulation.

Feedback trades raw gain for controlled behavior. With open-loop gain (A) and feedback factor (eta), the closed-loop gain is

$$A_{CL}=\frac{A}{1+A\beta}$$

When loop gain (A\beta) is large, closed-loop accuracy depends mainly on passive ratios, but finite gain leaves error. Feedback can reduce distortion and output impedance while extending useful bandwidth. It cannot remove input-referred noise or offset already inside the loop, and it can become unstable when phase lag approaches the condition for positive feedback.

Frequency compensation shapes loop gain so unity crossover retains adequate phase margin under load, process, temperature, supply, and common-mode variation. Dominant-pole compensation is robust but may sacrifice bandwidth. Miller compensation, feed-forward zeros, nested loops, and multistage amplifiers require explicit pole-zero analysis. Capacitive loads and package inductance are part of the loop. Stability should be checked for every meaningful operating state, including startup and output-current transitions.

Noise is a statistical signal budget. Resistors generate thermal noise; MOS channels produce thermal noise; traps create low-frequency flicker noise; junctions and currents can contribute shot noise. Each source is referred to a common point through its transfer function, and uncorrelated mean-square contributions are summed. For sources with RMS values (v_{n,i}),

$$v_{n,total}=\sqrt{\sum_i v_{n,i}^2}$$

Spectral density must be integrated over the actual noise bandwidth. A filter’s equivalent noise bandwidth is not always its nominal cutoff. Chopping and auto-zero techniques shift or sample offset and flicker noise, but introduce ripple, aliasing, switching artifacts, and charge injection. Larger input devices can reduce flicker noise and mismatch while increasing input capacitance.

Linearity describes deviation from an intended transfer function. Harmonic distortion, intermodulation, compression, integral nonlinearity, and differential nonlinearity reveal different failures. Transistor nonlinearities can be reduced through degeneration, feedback, differential symmetry, complementary paths, and calibration. Passive components also vary with voltage and temperature. A circuit that is linear for a small sine wave may behave differently for multi-tone or near-rail signals.

Differential signaling rejects disturbances common to both inputs and reduces even-order distortion, but only to the extent that paths match. Common-mode feedback sets output common mode in fully differential amplifiers and forms its own loop with stability requirements. Input common-mode range, output swing, and device stacking determine whether all transistors remain in their intended region.

Matching is local statistical control. Threshold voltage, mobility, resistor value, and capacitor density vary globally and locally. Common-centroid geometry, interdigitation, unit elements, identical orientation, dummy structures, and symmetric routing reduce gradient and edge effects. Increasing device area usually improves random matching approximately with inverse square root of area, but costs capacitance and can worsen speed.

Systematic mismatch often dominates after devices grow large: different drain voltage, well proximity, metal density, stress, temperature, or routing resistance. Layout-dependent effects must be included in extracted simulation. Trimming and calibration remove some residual error, but require observability, storage, test time, and a model of how error moves over temperature and life.

Data converters bridge continuous and discrete representations. An ideal (N)-bit ADC has an LSB near full-scale range divided by (2^N), but real performance is constrained by noise, distortion, clock jitter, comparator metastability, capacitor mismatch, reference settling, and digital coupling. Architectures choose different tradeoffs: flash is fast and large, SAR is efficient and versatile, pipeline supports high throughput with latency, and delta-sigma trades oversampling and filtering for resolution.

For an ideal full-scale sine wave, quantization-limited signal-to-noise ratio is approximately (6.02N+1.76) dB. Effective number of bits derived from measured SNDR is more informative than nominal code width. Aperture jitter limits high-frequency input performance because time uncertainty becomes voltage error proportional to input slew rate.

DACs use resistor strings, current steering, capacitor arrays, or hybrid structures. Static matching controls INL and DNL; timing mismatch and switch charge create dynamic glitch energy. Segmentation uses thermometer coding for the most significant elements and binary weighting elsewhere to balance area, decoding, monotonicity, and switching.

Power integrity is signal integrity for analog blocks. Supply ripple modulates bias, gain, oscillator frequency, and converter references. PSRR varies with frequency and operating point. Local regulation, filtering, differential circuits, careful return paths, and supply partitioning help, but no schematic symbol creates an ideal ground. Extracted resistance and inductance reveal shared impedance through which one block disturbs another.

Analog layout is circuit design expressed geometrically. Parasitic capacitance moves poles and slows settling; series resistance adds noise and gain error; coupling creates feedthrough; metal current density affects reliability. Sensitive high-impedance nodes are compact and shielded. Differential routes match length, layer, surroundings, and via count. High-current paths use wide metals and redundant vias. Devices that must match share orientation and environment.

Post-layout extraction is followed by simulations across process corners, mismatch Monte Carlo, supply, temperature, load, and parasitic variation. Corners cover bounded global shifts; Monte Carlo estimates local statistical spread. Neither substitutes for the other. Statistical tails need sufficient samples or analytical importance sampling, especially when the target failure probability is much lower than a routine simulation run can observe.

Measurement completes the model loop. Bench setups add source impedance, cable loss, probe capacitance, instrument noise, fixtures, grounding, and calibration error. De-embedding and correlation structures separate circuit behavior from setup behavior. Production tests use faster proxies than characterization, so correlation must show that those proxies catch relevant failures without excessive yield loss.

Silicon data should be viewed as distributions across die, wafer, lot, temperature, and time. A typical-unit plot can conceal systematic edge effects or bimodal populations. Failure analysis connects an electrical signature to layout, process, package, or model assumptions. The strongest teams feed that learning into device characterization, design checks, reusable IP, and test limits.

Professional analog design is controlled uncertainty. Start with a measurable signal budget, choose operating points from process data, use feedback with stability evidence, allocate noise and linearity by stage, plan matching in layout, simulate extracted parasitics and variation, and correlate against silicon. The finished block is not merely functional at a nominal schematic point; it retains useful information across the physical conditions in which the product must work.