← Back
Physical Intuition / Blog

Analog Design Bench

Measuring frontier coding agents on original, long-horizon analog circuit design tasks

Analog circuit design has always been the domain where intuition meets physics. We built Analog Design Bench to measure whether AI agents can navigate real design trade-offs: PVT corners, Monte Carlo yield, startup transients, noise budgets. Fifty tasks, twenty-one models, thousands of SPICE simulations.

Numbers
50Tasks
21Models Evaluated
79%Best Pass@1
Task Examples

Design a Sky130 high-impedance first-order bandgap core

Build a startup-safe CTAT/PTAT bias/reference core meeting voltage, tempco, line-regulation, supply-coupling, power, noise, and Monte Carlo limits across PVT.

bandgappvtmonte-carlo

Design a 1 GHz Sky130 Divide-by-Two Flip-Flop

Using the SKY130 130 nm CMOS process, design a transistor-level divide-by-two flip-flop operating at 1 GHz.

digitalhigh-speed

Design a Sky130 2.4 GHz Inductive-Degeneration LNA

Design a transistor-level 2.4 GHz inductively degenerated LNA with 12 dB gain and 2 dB noise figure.

rflnapvt

Design a Sky130 3-bit Flash ADC

A clocked 3-bit flash ADC graded on linearity, dynamic coding, delay, overdrive, and power across PVT.

adcmixed-signal

Design a 3.3-to-3.8 GHz broadband 50-to-200 ohm match

Synthesize a finite-Q lumped passive network that matches a 200 ohm RF load to a 50 ohm source.

rfmatching

Design a Sky130 capacitively coupled neural amplifier

Build a micro-power capacitively coupled neural amplifier meeting gain, band-corner, stability, noise, and distortion targets.

biomedicallow-power
Example Trace

A passing run from claude-fable-5 on the bandgap reference task. The agent wrote this SPICE netlist after 42 steps and 44 minutes, passing all 12 verification checks.

circuit.spi
* First-order SKY130 bandgap reference core
* Topology: PTAT core with Q1 (1x) vs Q2 (24x) + R1
* Output: mirrored PTAT current into R2 + Q3 gives vref ~ 1.22V

.subckt bandgap_reference vss vdd vref

* PTAT core BJTs
XQ1 na na vss vss sky130_fd_pr__npn_05v5_W1p00L1p00 mult=1
XQ2 nq2 nq2 vss vss sky130_fd_pr__npn_05v5_W1p00L1p00 mult=24
R1 nb nq2 17k

* Output branch
XQ3 nc nc vss vss sky130_fd_pr__npn_05v5_W1p00L1p00 mult=1
R2 vref nc 92k
CO vref vss 10p

* PMOS mirror
XMP1 na pgate vdd vdd sky130_fd_pr__pfet_01v8_lvt l=8 w=32 nf=4
XMP2 nb pgate vdd vdd sky130_fd_pr__pfet_01v8_lvt l=8 w=32 nf=4
XMP3 vref g3 vdd vdd sky130_fd_pr__pfet_01v8_lvt l=8 w=32 nf=4

* Error amplifier (5T OTA)
XM1 x1 nb ntail vss sky130_fd_pr__nfet_01v8_lvt l=1 w=16 nf=4
XM2 pgate na ntail vss sky130_fd_pr__nfet_01v8_lvt l=1 w=16 nf=4
XML1 x1 x1 vdd vdd sky130_fd_pr__pfet_01v8_lvt l=8 w=32 nf=4
XML2 pgate x1 vdd vdd sky130_fd_pr__pfet_01v8_lvt l=8 w=32 nf=4

.ends bandgap_reference
Verifier Results9/9 PASS
PVT reference voltage1.208–1.228 V
PVT power78.1 µW max
Temperature coefficient24.1 ppm/°C max
Line regulation5.2 mV/V max
Startup settling< 10 µs
Startup overshoot0.0 mV
Supply rejection0.067 V/V max
Integrated noise119 µVrms
Monte Carlo yield100% (30 seeds)
Leaderboard (Top 12)
#ModelPass@1CostOutputTime
1claude-fable-5[max]79%$21.88113k58m
2gpt-5-6-sol[max]67%$13.9175k52m
3claude-opus-5-max[max]66%$26.37177k1h 41m
4claude-opus-4-8[max]63%$18.52199k1h 18m
5gpt-5-6-sol[xhigh]57%$7.0048k36m
6kimi-k3-1m54%$3.81103k1h 34m
7gpt-5-6-terra[max]48%$8.15101k1h 04m
8qwen-3-8-max45%$4.00206k1h 51m
9deepseek-v4-flash-073143%$0.18384k1h 21m
10glm-5-242%$16.48174k1h 41m
11doubao-seed-evolving41%$1.50121k1h 37m
12gpt-5-6-luna[max]41%$1.03120k1h 12m
Methodology

Each agent receives a natural-language spec, a starter netlist, and access to ngspice. It must iterate on the circuit until all verification checks pass. No hints, no partial credit. The evaluator runs the same SPICE testbenches used to sign off real silicon.