modelnoise_examples

Production compact-model noise validation — Enhancement-165

Enhancements 159–161 validated a real compact model's DC, coverage, and small-signal (AC / C-V / fT) behavior; E-164 its large-signal RF. This one exercises the remaining untested small-signal path — OSDI's .noise stamping of the models' own white_noise / flicker_noise sources — on production models, compiled in place from the OpenVAF integration-test sources.

BSIM4 — validated against ngspice's built-in

The output-noise spectral density Sv(f) of a BSIM4 common-source amplifier is compared to ngspice's built-in BSIM4 over the full band (Panel A). Both the low-frequency 1/f flicker region (Sv ∝ 1/√f) and the flat high-frequency thermal floor match to < 2 % — a stringent check that the OSDI noise stamping reproduces the native model's noise physics.

openvaf-r ../../OpenVAF-master-20260610/integration_tests/BSIM4/bsim4.va -o bsim4.osdi
ngspice -b modelnoise_demo.cir

HICUM — validated against shot-noise physics

HICUM/L2 (SiGe HBT) has no ngspice built-in, so its noise is checked against physics. Its default noise is white (shot + resistance thermal, no flicker by default) — a flat spectrum, in clear contrast to the MOSFET's 1/f rise. With a small source resistance so the intrinsic device noise dominates, the output-noise floor tracks the collector shot-noise line √(2q·Ic·RC²) across two decades of bias current (Panel B) and scales as √Ic — the defining bipolar white-noise behavior.

Verify

python3 verify_modelnoise.py    # 5 checks, under BOTH the Sparse and KLU solvers
python3 make_modelnoise_fig.py  # -> modelnoise.png

• [1] OSDI BSIM4 output-noise spectrum matches built-in BSIM4 to < 4 % (≈ 1.5 %).
• [2] BSIM4 shows the 1/f flicker region (Sv(1Hz)/Sv(10Hz) ≈ √10).
• [3] BSIM4 shows the flat thermal floor at high frequency.
• [4] HICUM noise is white (flat) at mid-band — no flicker by default.
• [5] HICUM output floor tracks the collector shot noise 2q·Ic·RC² and scales as √Ic.

.noise runs under both solvers (KLU included, since E-113 fixed the KLU adjoint solve).

Why the results are physically correct

• Flicker (1/f). Trap-related carrier-number fluctuation gives a power density ∝ 1/f, so the amplitude density falls as 1/√f — exactly the low-frequency slope, matching the built-in model.
• Thermal floor. Channel/resistance thermal noise is white, giving the flat high-frequency floor.
• Shot noise. A DC current Ic crossing a junction carries shot noise i² = 2q·Ic; through the RC load that is 2q·Ic·RC² at the output, tracking the collector current across the bias sweep.

See Enhancement-165 for the full write-up.