Content is user-generated and unverified.

Cascode LNA — Head-to-Head Comparison of Three MNA Implementations

TSMC N7 FinFET · 3.5 GHz n78 · Browser-Based Interactive RF IC Tools


1. Overview

Three single-file HTML/Three.js implementations of an inductively-degenerated cascode LNA, each built on a Modified Nodal Analysis (MNA) engine but committing to a different design priority:

IDShort nameHeadline priority
A1a3e80966Multi-RHSCleanest MNA engineering; multi-RHS noise solver
A22cdb351cDE + Tapped-CRFIC-complete; output matching network + DE optimizer
A3dbb6cd1aCorrelated-Noise v3.2Highest physical fidelity; van der Ziel correlated noise

All three share the same die-view skeleton (24 × 18 µm die canvas, M1–M9 stack visualization, six bondpads, animated signal-flow particles, X-ray / rotate / reset controls, debounced rebuild on parameter change).


2. Detailed Comparison Matrix

2.1 Circuit Topology

AxisA1 (Multi-RHS)A2 (DE + Tapped-C)A3 (Correlated-Noise)
MNA size6 nodes6 nodes5 nodes (pad split from gate)
Node labelsnin, ng1, ns1, nd1, nd2, noutsamepad, gate, src, cas, out
Input pad modelC_pad = 200 fF lumpedC_pad = 200 fF lumpedC_bond = 30 fF + C_esd = 170 fF split
Output pad capC_out = 100 fF
DC-block Cd10 pF ✓10 pF ✓
Output matchingnone (direct tap)Tapped-C (C1 series + C2 shunt)none
RL load on output

2.2 Noise Modeling

AxisA1A2A3
Total noise sources9611
Included sourcesRs, Rg+R_Lg, R_Ls, M1 ch, M2 ch, rds1, rds2, Rp_Ld, RLRs, Rg, R_Ls, M1 ch, M2 ch, R_LdRs, Rg, M1 ch + iG + correlation, M2 ch, rds1, rds2, R_Ls, R_Ld, substrate
MOSFET channel γ0.67 (long-ch)0.671.2 (short-ch FinFET) ✓
Induced gate noise (δ)✓ δ = 3.5
Channel ↔ gate correlation✓ van der Ziel j·|c|·√(S_id·S_ig), |c| = 0.395
Substrate noiseR_sub = 200 Ω, C_sub = 50 fF
Noise solve methodMulti-RHS — single Gauss-Jordan factorization solves signal + all 9 noise vectors ⚡Per-source solveY × 6Per-source csolve × 10+

2.3 Optimization

AxisA1A2A3
OptimizerGD + A* (selectable)DE / rand / 1 / bin onlyGD + A* (selectable)
Parameter space5 (W, Id, Lg, Ls, Ld)5 DE + 2 slider (W, Id, Ld, C1, C2 + Lg, Ls)5 (W, Id, Lg, Ls, Ld)
DE population × gens35 × 90
GD featuresmomentum (0.85), adaptive LR, stagnation kickmomentum, adaptive LR, stagnation kick
A* featuresdiscrete grid expansion, closed setsame

2.4 Output Metrics & HUD

MetricA1A2A3
Gain
NF
S₁₁
S₂₂
IIP3
Power
g_m
f_T
Rollett K
NF breakdown bars (Rs / Rg / M1 / M2 / Ld)
Node voltage magnitudes (gate / cas / out)

2.5 Specifications

SpecA1 (min / ideal / max)A2A3
Gain15 / 18 / 25 dB15 / 18 / 20 dB (tightest)15 / 18 / 25 dB
NF0.5 / 1.5 / 3.0 dB0.5 / 0.8 / 1.5 dB (most realistic n78)0.8 / 1.2 / 2.0 dB
S₁₁−25 / −15 / −10 dB−25 / −15 / −10 dB−25 / −15 / −10 dB
S₂₂−20 / −12 / −10 dB
IIP3−5 / 0 / 10 dBm−5 / 2 / 5 dBm−5 / 0 / 10 dBm
Power2 / 5 / 12 mW5 / 8 / 15 mW (industrial range)2 / 5 / 12 mW

2.6 Visualization

FeatureA1A2A3
Spiral inductors (Lg / Ls / Ld)
Tapped-C 3D rendering (C1 / C2 stacks) (scales with C1/C2 value)
Signal particle color = NF✗ (speed only)
MIM cap MIMO oxide layers
Component labels with live values✓ (+ C1, C2 labels)

3. Where Each Implementation Wins

A1 — Multi-RHS is the cleanest engineering

The multi-RHS solver is the textbook-correct way to do MNA noise analysis: you factor the admittance matrix Y once, then back-substitute every noise excitation vector (I_n1, I_n2, ...) against the same triangulation. SPICE does this. For EL703r, A1 is the version that shows students how a real noise simulator is structured under the hood. It accepts the long-channel γ = 2/3 and skips induced-gate noise, so it under-predicts NF by roughly 1–1.5 dB on N7 — but the machinery is the most pedagogically honest.

Best for: teaching MNA mechanics; benchmarking solver efficiency.

A2 — DE + Tapped-C is the most RFIC-complete

A2 is the only version that closes the output matching loop with a tapped-C network (C1 series + C2 shunt to ground), reports S₂₂, and reflects the C-network in 3D geometry that scales visually with C1 / C2 values. Differential Evolution is the right optimizer choice for the resulting seven-dimensional space: GD and A* would stall on the C1 ↔ C2 ridge, while DE/rand/1/bin with F = 0.8, CR = 0.9 traverses it cleanly. The spec table (15–20 dB gain, 0.8–1.5 dB NF, 5–15 mW) is also the most honest n78 LNA target of the three. Trade-off: only 6 noise sources, no correlation.

Best for: teaching the full RF design flow (input + output matching, two-port S-params, evolutionary optimization).

A3 — Correlated-Noise v3.2 has the highest physical fidelity

A3 is the only version with:

  • γ = 1.2 (correct for short-channel FinFETs — long-channel 2/3 underestimates N7 NF significantly)
  • Induced gate noise S_ig = 4kT·δ·(ωCgs)² / (5·g_m)
  • The j·|c|·√(S_id·S_ig) correlation term that lets channel and gate noise contributions partially cancel through Vn_M1[4]*· Vn_IG[4]
  • Cbond / Cesd split (the lumped 200 fF over-isolates the gate from the pad)
  • Substrate coupling through R_sub / C_sub
  • Rollett K-factor, so you can see when the optimizer wanders into conditional stability

Trade-off: 10+ separate factorizations per fitness evaluation. The optimizer feels it.

Best for: teaching noise physics; research-grade NF modeling; preparing publication-quality LNA simulations.


4. Recommended v4 Merge

Borrow the best of each:

FromTake
A1Multi-RHS solver (one factorization, all noise vectors)
A3γ = 1.2, induced gate noise, j·|c|·√(S_id·S_ig) correlation, Cbond / Cesd split, substrate, Rollett K
A2Tapped-C output network, S₂₂ in HUD, DE optimizer for the 7-D space, NF breakdown bars

The result is the version worth handing to an EL703r student as a reference implementation — and the natural starting point if any of these is going to support a paper on noise-aware LNA optimization or feed into the EM × λ-aware P&R flow (parasitic noise routing).


5. Summary at a Glance

PriorityChoose
Solver efficiency / pedagogy on MNA mechanicsA1
Complete RFIC two-port design flowA2
Research-grade noise modelingA3
EL703r teaching artifactA1 for solver, then A3 for noise
Conference-paper-grade simulationA3 → merge with A1's solver

Comparison prepared for Yao-Jen Chang · Intelligent Assistive Technology Lab · CYCU · EL703r Health-Care Technology Topics I

Content is user-generated and unverified.
    Cascode LNA Comparison: 3 MNA Implementations Analyzed | Claude