Iron, Glass, and Orange Light: A 1950s-Style Adder Built From Passive Parts
A build guide for a 4-bit visible adder using only resistors, capacitors, inductors, transformers, and neon bulbs. No transistors, no integrated circuits, no semiconductors of any kind — every active element is a piece of glowing gas, every controlled element is a piece of wound iron. The technology was buildable by the mid-1940s and would have been recognisable to anyone who worked on SAGE or the Soviet Setun. It is buildable today for under $200 in a maker space with a CO₂ laser cutter and a patient weekend of transformer winding.
The 1947 Bell Labs press conference announcing the point-contact transistor hadn’t happened yet. The engineers at MIT Lincoln Lab designing SAGE, the engineers at Penza Plant 50 designing the Setun, and the engineers at the Royal Radar Establishment designing the early ramark systems were all solving the same problem: how do you build digital logic without any active semiconductor element?
The answer they converged on — independently, on three continents — was magnetic amplifier logic, or mag-amp logic: saturating transformers with multiple windings, used as the active switching element. With appropriate biasing and a sinusoidal supply, a single small ferrite-cored transformer is a complete logic gate. Combined with neon bulbs as visible state indicators and bistable storage elements, the whole logic family falls out.
This post is the build guide. The end state is a small wooden-framed device about the size of a hardcover book, with twenty-some glowing orange neon bulbs visible across the front, an audible 60 Hz hum from the transformer in the base, and a row of brass-handled toggle switches that the operator (you, or a kid) uses to enter two 4-bit numbers. Press a “compute” lever and the answer appears as a binary pattern across the output bulbs.
Every active element glows. Every switching element is wound copper on iron. No silicon anywhere. The whole machine is buildable using techniques from 1948.
1. The constraint, and why it matters
If you’re allowed transistors and ICs, building visible adders is a solved problem — drop a 74HC283 four-bit adder on a board, wire its inputs and outputs to neon bulbs through driver transistors, done. The machine you build looks impressive but it’s a cheat: the computation happens invisibly inside a black silicon package, and the bulbs are just decoration.
If you take away the active semiconductors, the design problem becomes genuinely interesting. You have:
- Resistors — passive voltage division, current limiting, thermal load
- Capacitors — passive energy storage, AC coupling, time delays
- Inductors and transformers — passive energy storage, voltage transformation, magnetic coupling between circuits
- Neon bulbs — the only nonlinear element you’re allowed
The neon bulb is the entire reason the system works. Below about 90 V it’s an open circuit (insulator). Above 90 V it ignites and conducts at about 60 V drop with about 2 mA of current. Once lit, it stays lit until the voltage across it drops below 60 V. This hysteretic, current-limited nonlinearity is what makes logic possible.
Everything else in the circuit is linear, and linear circuits can’t make decisions. The neon bulb is where the decisions happen.
2. The physics, in 400 words
A magnetic amplifier is a transformer with a saturable ferrite core and multiple windings. The trick is that a ferrite core’s permeability collapses dramatically when the core is driven into saturation — typically by an order of magnitude or more. Two regimes:
- Below saturation: core has high permeability, the gate winding has high inductance, AC current through it is choked off, very little current reaches the load.
- Above saturation: core has near-vacuum permeability, the gate winding has nearly zero inductance, AC current through it flows freely, full current reaches the load.
You arrange the windings so that control currents push the core toward or away from saturation. Each control winding contributes flux proportional to its current. The sum of fluxes either pushes the core over the saturation threshold (output ON) or doesn’t (output OFF). The control inputs are now the inputs of a logic gate.
For neon-based logic, the load is a neon bulb in series with a current-limiting resistor. When the core is unsaturated, the gate winding chokes off the AC supply current and the neon doesn’t see enough current to fire (or maintain its sustaining voltage between half-cycles). When the core is saturated, the gate winding passes the AC freely, the neon fires each cycle, and the eye sees a steady orange glow (because the 60 Hz cycle is faster than visual persistence).
The control windings carry DC bias derived from upstream stages. The DC comes from earlier neon bulbs being lit or dark — the current flowing through a lit bulb is the “1” signal, the absence of current is the “0” signal.
So the architecture is:
AC supply → saturable transformer (with DC control windings from inputs) → neon bulb → DC output to next stage
Each stage’s neon is both the output indicator (visible orange glow) and the source of DC bias for the next stage’s control winding. The 60 Hz AC supply provides the natural reset every cycle. The whole logic family emerges from this one architectural pattern.
3. The gate: a single mag-amp cell in detail
Here’s the canonical mag-amp gate, drawn out:
AC supply (100 V RMS, 60 Hz)
│
▼
┌───────┐
│ Gate │ ←──┐
│ wind. │ │ Control winding 1 (input A)
│ N=200 │ ←──┤ Control winding 2 (input B)
│ │ │ Reset winding (counter-flux, DC bias)
│ ferr. │ ←──┘
│ core │
└───┬───┘
│
▼
┌────────────────┐
│ R_limit ~10 kΩ │
└────────┬───────┘
│
▼
╭─────────╮
│ NE-2 │ ← visible output bulb
╰────┬────╯
│
▼
─┴─ ground (or DC bias return)
│
▼
output current
(feeds next stage's
control winding)
The core is a small ferrite toroid — Magnetics 0R45614TC is a workable choice, or a Fair-Rite 5961002701 in the same size range. Outer diameter about 14 mm, inner diameter about 8 mm, height 5 mm. Saturation flux density about 0.3 T, residual flux density about 0.1 T.
The gate winding is 200 turns of 32 AWG (0.2 mm) magnet wire. Each control winding is 50 turns. The “reset” winding (which counter-biases the core, holding it below saturation when no input is asserted) is 100 turns and is fed by a small bias resistor from the DC rail.
Wound up, the transformer is about the size of a fingernail. Five minutes of winding per gate, once you have the jig.
The output bulb is a generic NE-2 neon indicator. The current-limit resistor sets the steady current to about 1.5 mA when lit. The bulb glows continuously when the core is saturated (each AC half-cycle drives enough current through the neon to keep it ignited; the next half-cycle doesn’t quite extinguish it because of the dwell time of plasma deionization).
4. The basic gates, all from one trick
Different gate functions come from different winding configurations on the same basic transformer. The shared mechanism: control windings contribute flux toward or away from saturation, the threshold is set by the reset winding’s counter-bias.
AND
Two control windings of 50 turns each, both wound in the same flux direction. The reset winding is biased so the core just barely fails to saturate when one control input alone is asserted. When both control inputs are asserted, their combined flux pushes the core into saturation. Output neon fires only when both inputs are present.
OR
Same construction as AND but with the reset bias relaxed. Now one control input alone is enough to push the core into saturation. Either input alone fires the output.
NOT (inverter)
The control winding is wound in opposition to the gate winding’s flux direction. The reset winding holds the core close to saturation by default. Input current in the opposition-wound control winding removes flux from the core, pulling it away from saturation. Output neon is normally lit (input absent → core saturated → AC passes → neon fires) and dark when input is present.
NAND, NOR
AND or OR followed by NOT — but you can do this with one transformer by adding a second polarity-reversed control winding. A NAND gate is the AND configuration plus an extra reset winding driven by a constant DC bias such that the output polarity is inverted relative to AND. One transformer per NAND gate.
XOR — the hard one
There’s no single-transformer XOR configuration that works cleanly. Two practical builds:
Build A (transformer-bridge): Two mag-amp transformers wired such that the output neon sits between their two secondaries. Each transformer is configured as an asymmetric AND. The two ANDs feed the output bulb in opposite polarity. Either input alone causes one or the other AND to saturate and fire the output; both inputs cause both ANDs to fire, cancelling each other at the output node. Two transformers and one output bulb per XOR.
Build B (NAND construction): XOR(A, B) = NAND(NAND(A, NAND(A,B)), NAND(B, NAND(A,B))). Four NANDs from above. Four transformers and one output bulb per XOR, but uses the same gate topology four times — easier to fabricate identical parts.
Build A is more elegant; Build B is more “textbook computer architecture.” For a hobby build I’d pick Build B because the parts are identical and debugging is easier.
5. The supply
A linear AC supply derived from a small step-down line transformer is the right approach. You’re not allowed any active rectification or regulation, so the supply is exactly as simple as possible: line voltage in, step-down transformer, RMS output around 100 V at maybe 50 mA capacity.
Specific construction:
- Primary: 120 V mains (or 240 V depending on region), via a fused inlet and a power switch.
- Isolation transformer: an off-the-shelf 1:1 isolation transformer rated for 100 VA is the right first stage. Galvanic isolation from mains is non-negotiable for safety. Cost: $40-60.
- Step-down transformer: secondary of the isolation feeds the primary of a step-down with a 120:100 ratio (so a small step-down, basically just to clean up the waveform and give you a clean sinusoidal output at the operating voltage you want). Or use a single 120:100 isolation transformer if you can find one.
- Tank capacitor for waveform cleanup: a 1 μF / 250 V film cap across the secondary suppresses any switching noise from the wall.
The secondary feeds the gate windings of every mag-amp transformer in the device. Total current draw with all 20+ bulbs lit is around 40 mA.
A small bias DC rail for the reset windings comes from a half-wave rectified branch of the secondary — but wait, no diodes either if you’re truly strict. Two options:
- Allow a single Selenium rectifier stack (pre-1948 technology, no silicon involved). Cost ~$15 from electronic surplus dealers, looks gorgeous, lasts forever.
- Pure-passive bias via resonant LC tank tuned to the second harmonic. A simple bias network using only L and C produces a DC offset by exploiting the nonlinearity of the neon bulbs in the bias branch themselves. More elegant, harder to design.
For the build guide I’ll specify Selenium rectifiers, because they predate the transistor era by decades and the 1900s constraint allows them.
6. The 4-bit adder, fully specified
A 4-bit ripple-carry adder needs four full-adder cells. Each full-adder cell computes:
SUM = A XOR B XOR C_in
CARRY = (A AND B) OR (C_in AND (A XOR B))
Per-bit component count using mag-amp logic with Build B XORs:
| Gate | Mag-amp transformers | Neon bulbs |
|---|
| XOR(A, B) — 4 NANDs | 4 | 4 |
| XOR(prev, C_in) — 4 NANDs | 4 | 4 |
| AND(A, B) | 1 | 1 |
| AND(XOR_AB, C_in) | 1 | 1 |
| OR (carry parts) | 1 | 1 |
| Per full-adder cell | 11 | 11 |
For 4 bits, that’s 44 transformers and 44 output bulbs in the adder core, plus:
- 4 + 4 = 8 input register bulbs (one per input bit of A and B), driven by toggle switches with current-limit resistors in series
- 4 sum-output bulbs (visible answer)
- 1 carry-out bulb (overflow indicator)
- 1 “compute” indicator (lit while computation is active, to give the kid a sense that the machine is “thinking”)
Total visible bulbs: 58. Total mag-amp transformers: 44. Total passive components: about 200 resistors and capacitors.
This is a substantial build. About a long weekend of transformer winding once the cores arrive, plus another weekend of PCB assembly and tuning. The result is, I think, the most beautiful pure-passive computer you can build at hobby scale. Every wire that carries information is visible. Every active element glows orange. Every control signal is a current through an iron core that you wound yourself.
7. Build plan
Bill of materials
| Item | Qty | Spec | Source | Cost |
|---|
| Ferrite toroid core | 50 | 14 mm OD, FairRite 5961002701 or equivalent | Mouser, eBay | $0.80 each |
| Magnet wire | 1 reel | 32 AWG (0.2 mm), 500 ft | Digi-Key, Amazon | $30 |
| NE-2 neon bulb | 60 | Generic indicator | Mouser, Amazon | $0.30 each |
| 1/4 W resistors | 100 | Assorted values (1 kΩ to 1 MΩ) | Mouser kit | $20 |
| Film capacitors | 50 | 0.1 μF, 250 V | Mouser | $0.30 each |
| Toggle switches (DPDT) | 12 | Brass handle, panel-mount | Adafruit, eBay | $4 each |
| Wooden enclosure | 1 | 30 × 20 × 8 cm hardwood | Local woodshop, scrap | $30 |
| Brass faceplate | 1 | Laser-cut 20 × 30 cm | Online laser service | $25 |
| Isolation transformer | 1 | 100 VA, 1:1 | eBay, surplus | $50 |
| Step-down transformer | 1 | 120:100 V, 100 VA | Surplus | $40 |
| Selenium rectifier stack | 1 | Vintage, ~250 V at 50 mA | eBay | $15 |
| Fuse holder + fuse | 1 | 250 V, 250 mA slow-blow | Mouser | $5 |
| Power inlet (IEC C14) | 1 | With switch and fuse | Mouser | $8 |
| Hookup wire | 50 ft | 22 AWG stranded, multiple colors | Adafruit | $20 |
| PCB | 1 | Two-layer, 200 × 250 mm | JLCPCB or OSHPark | $30 |
Total component cost: approximately $190.
| Tool | What for | Cost if buying |
|---|
| Soldering iron | Through-hole assembly | Already have |
| Winding jig (DIY) | Holding the toroid while you wind | Build from scrap wood and a small hand-drill chuck |
| Multimeter | Voltage testing | Already have |
| Variac (variable autotransformer) | Bringing up the supply gradually during first power-on | Borrow from maker space, or do without |
| Oscilloscope | Watching gate waveforms during tuning | Borrow, or do without (visual inspection of bulb states works) |
| Hand drill | Wooden enclosure construction | Already have |
| Wood-finishing supplies | If you want the steampunk aesthetic | $30 |
This is the bulk of the build. Each transformer needs:
- Gate winding: 200 turns of 32 AWG on the toroid. Start with a 30 cm tail, wrap 200 turns evenly distributed around the core, end with a 30 cm tail. The tails become the connections to the rest of the circuit.
- Control windings: 50 turns each, in the right flux direction for the gate’s logical function. Mark the direction carefully — winding direction determines whether a control input drives the core toward or away from saturation. I use a Sharpie dot on the core to indicate “start” direction.
- Reset winding: 100 turns, wound to oppose the control windings (so the reset bias keeps the core below saturation when no inputs are asserted). Fed by a fixed bias resistor from the DC bias rail.
A winding jig is a small handheld drill with a chuck that grips a wooden form holding the toroid. Crank the drill slowly while feeding wire. With practice you can wind a 200-turn gate winding in about 3 minutes.
Wind all 50 transformers before starting the PCB work — it’s monotonous but goes faster if you batch it.
PCB and assembly
Design the PCB in KiCad with the layout matching the logical structure of the adder. Each full-adder cell occupies a roughly 60 × 60 mm region. The transformers mount in through-holes, the neon bulbs sit in dedicated holes that come up through the brass faceplate, the resistors and caps lay flat on the board.
The aesthetic move: mount the bulbs on the underside of the brass faceplate, with the wood frame holding everything together. From the front, you see polished brass with orange-glowing holes. From the back, you see neat rows of wound iron toroids on a PCB. Both sides are photogenic.
Assembly order:
- Build the supply section first. Verify clean ~100 V AC and ~120 V DC bias before installing anything downstream.
- Build and test one full-adder cell. Verify all 8 input combinations produce the correct output. Tune the reset bias to set the threshold right.
- Build the remaining three cells, reusing the tested topology.
- Wire the ripple-carry path.
- Wire the input toggles and output indicator bulbs.
- Final assembly into the wooden case.
8. Safety
100 V AC from a transformer secondary, while much less dangerous than mains AC, is still capable of causing painful shocks and (under unusual circumstances) lethal cardiac fibrillation. Take it seriously.
Non-negotiable rules:
- Isolation transformer between mains and everything else. No exceptions. The whole device’s secondary side floats relative to earth ground. A single hand-to-device contact is then just a one-point contact and won’t pass current.
- Fuse the primary. A 250 mA slow-blow fuse in the IEC inlet protects against catastrophic failure (a shorted transformer winding could otherwise melt the supply harness and start a fire).
- Enclose all live conductors. The wooden case has the only user-accessible surface as the front brass plate with its toggles and bulb holes. Everything behind it should be screwed shut, not snap-fit. A determined kid with a screwdriver could still get in, but they’d have to mean it.
- One-handed operation rule. Anyone working on the powered device keeps one hand in their pocket. Current through the chest is the dangerous path; one-handed contact bypasses the chest entirely.
- Bleeder resistors across any DC capacitor. Even with no semiconductor regulation, the bias capacitor can hold a 100 V charge for minutes after disconnect. A 1 MΩ resistor across each cap drains it in seconds.
- Warning label. Inside the case, a label reading “100 V AC internally. Service by qualified personnel only. Disconnect mains and allow 30 seconds before opening.”
If your maker space has a hard cap at “low voltage only,” this project is over the line. Talk to leadership before starting.
9. What you’d actually see
Power on. The supply takes a couple of seconds to come up; the 60 Hz hum from the transformer is audible. The “compute” indicator bulb is dark (no compute happening). The input register bulbs reflect whatever the toggle switches are set to — flip a switch, the corresponding bulb lights orange or goes dark.
Set A = 0101 (5) and B = 0011 (3) using the toggles. Eight bulbs across the input rows are lit in the appropriate pattern.
Press the “compute” lever. The compute indicator lights. Across the adder core, you see a cascade of bulbs lighting and extinguishing as the gates resolve. The XOR cells flicker as they settle. The AND cells either light or stay dark. The OR cells in the carry chain light in sequence as the carry ripples from bit 0 to bit 1 to bit 2 to bit 3.
The whole resolution takes about 200 ms — slow enough to watch. The 60 Hz supply means each gate has to wait at least one half-cycle to resolve, and the cascade through four ripple stages adds up.
When it settles, the sum bulbs show 1000 (8) and the carry-out bulb stays dark. The compute indicator turns off after a second to indicate “done.”
Press reset. All bulbs go dark. Set new inputs. Press compute again.
This is what computation looks like when computation is slow enough to see. No abstractions. No hidden silicon. Every signal that participates in the answer is a bulb you can point at.
10. Lineage and why it matters
The technology described above existed by 1947, was used heavily in the 1950s, and lost its battle with the transistor by 1960. Specific historical artifacts:
- The Soviet Setun ternary computer (1958) used mag-amp logic extensively. It is the most famous of the post-vacuum-tube, pre-semiconductor computers and operated until 1965. A working rebuild exists at Moscow State University.
- The SAGE air-defense system (deployed 1958-1983) used mag-amp logic for its signal-conditioning and parts of its display electronics. The IBM AN/FSQ-7 mainframe at the heart of each SAGE site was vacuum-tube-based, but the peripherals around it were heavily mag-amped.
- Several US Navy fire-control computers of the 1940s and 1950s (notably the Mk 56 and Mk 1A gun-director computers) used mag-amp logic combined with electromechanical resolvers. Many of these survived into the 1980s aboard cruisers and destroyers because they were genuinely irreplaceable — no transistor equivalent had been built.
- The Bendix G-15 personal computer (1956) used a hybrid of vacuum tubes and mag-amp logic.
These machines have been forgotten in the standard “valve → transistor → IC” narrative of computing history. They were a real technology, genuinely competitive with transistors for the better part of a decade, and lost not because they were bad but because semiconductor fabrication scaled exponentially while transformer winding did not.
What you’re building is a small descendant of those machines. It uses the same principles, the same kinds of components, and produces the same kind of visible orange-glow output. It is, in its small way, a working museum piece — a thing that demonstrates a road computing history could have taken further than it did.
It is also, I think, the most beautiful homemade computer you can build without a fab. Every signal glows. Every gate is wound iron. Every wire is doing work you can see.
If a child watches it count, the lesson is the same as for the pneumatic version, but with a different aesthetic. The pneumatic version says “this is computation as moving things.” The mag-amp version says “this is computation as iron, glass, and light.” They’re both right.
11. Going further
A few directions if you finish the basic adder and want more:
- Build it twice — two identical 4-bit adders side by side, wired so the carry-out of one feeds the carry-in of the other, gives you an 8-bit adder. Same components, double the bulbs.
- Add a multiplier. A 4-bit × 4-bit multiplier built from shift registers and conditional adds, all in mag-amp + neon, would be the most ambitious pure-passive computer build I’m aware of in current hobbyist work.
- Add an audio output. The 60 Hz hum from the transformer and the 60 Hz cycling of the gate windings make a natural musical signal. A small speaker coupled to the bias rail through a capacitor would produce an audible “computation hum” that shifts pitch slightly when different gates are active. Aesthetic bonus.
- Mount it in a display case with a small explanatory placard. The whole device is a working sculpture; treat it as one. A glass-fronted oak cabinet, brass faceplate engraved with operation instructions, a small typed history of mag-amp computing on a card beside it.
Coda
The thing I love about pure-passive logic is that it can’t be opened up to reveal a layer of cheating. There is no integrated circuit hidden behind the bulbs. There is no microcontroller in the base secretly doing the work. The wound iron transformers ARE the gates. The neon bulbs ARE the storage. The wire is just wire.
A modern computer is mostly an exercise in trust. You trust that the billion transistors inside the SoC are doing what the datasheet says they’re doing. You can’t see them. You can’t probe them. You can’t verify them. The machine works because the abstractions hold all the way down.
A mag-amp neon adder is the opposite. Every part of it is visible. Every signal can be probed with a meter. Every gate’s behaviour can be watched. The whole computation is in plain sight, in plain physics, running slowly enough to think about.
I want one. If you build one, send me a photograph. If you make a kit and put it on Tindie, I’ll buy one. The space of working visible computers is small; the space of working visible passive computers is smaller still. There’s room.
Further reading
- Magnetic Amplifiers — Theory and Application, William A. Geyger, McGraw-Hill, 1957. The textbook.
- Saturable Reactor Computers, R. C. Booton et al., Proceedings of the IRE, March 1957. Specific to mag-amp logic.
- Brusentsov, N. P. Setun: A Ternary Computer, Moscow State University, 1959. (Russian; partial English translations exist online.)
- MIT Lincoln Laboratory Journal, Vol 22 No 2 (2017): historical retrospective on SAGE.
- Neon Lamp Manual, General Electric, 1966. Chapter on relaxation oscillators and bistable circuits.
- Loebner, E. E. Cold Cathode Glow Discharge Tubes in Digital Computers, RCA Engineer, Vol 3 No 2, 1957.
- The Electronics of Selenium Rectifiers, M. R. Currie, Proceedings of the IRE, April 1947. For sourcing the bias supply.
- The Computer History Museum’s collection on pre-transistor computing has working mag-amp components on display.
- The Story of the SAGE Air Defense System, MITRE Corporation, 2005. Open-access PDF.
Comments, refutations, build photos, and corrections welcome. Especially refutations of the transformer winding counts — I haven’t prototyped this yet and the numbers may shift by 10-20% once a real build is on the bench.
![[photo of my neon clock]](https://web.archive.org/web/20170824164029im_/https://wwwhome.ewi.utwente.nl/~ptdeboer/ham/neonclock/neonclock.jpg)
![[schematic from Dance's book]](https://web.archive.org/web/20170824164029im_/https://wwwhome.ewi.utwente.nl/~ptdeboer/ham/neonclock/dance.jpg)
![[photo of one ringcounter]](https://web.archive.org/web/20170824164029im_/https://wwwhome.ewi.utwente.nl/~ptdeboer/ham/neonclock/ringcounter.jpg)
![[photo of one ringcounter with LDRs]](https://web.archive.org/web/20170824164029im_/https://wwwhome.ewi.utwente.nl/~ptdeboer/ham/neonclock/ringcounterldr.jpg)
![[photo of one ringcounter with LDRs]](https://web.archive.org/web/20170824164029im_/https://wwwhome.ewi.utwente.nl/~ptdeboer/ham/neonclock/ldr.jpg)
![[photo of clock with 2 blue leds]](https://web.archive.org/web/20170824164029im_/https://wwwhome.ewi.utwente.nl/~ptdeboer/ham/neonclock/blue.jpg)
