Moscow, 1958
Setun was built at Moscow State University by a team led by Nikolai Brusentsov, with the mathematician Sergei Sobolev behind the decision to attempt it at all. The first machine ran in 1958. It was a general-purpose computer, it was used for real work — teaching, engineering calculation, and later industrial control — and around fifty were produced. A successor, Setun-70, followed in 1970 with a stack architecture.
Its distinguishing property is the one this site is about: Setun computed in balanced ternary. Its word was 18 trits. Its arithmetic needed no two's complement, its comparison had three outcomes, and negation was a sign flip rather than an invert-and-add. Everything that balanced ternary is supposed to be good at, Setun did, on hardware, in production, sixty-eight years ago.
It is the standing counter-example to the claim that ternary computing is speculative. The speculative part has never been the arithmetic.
What Setun Was Actually Made Of
The element base is the whole story. Setun's logic was not transistors. It was ferrite cores and diodes, working as controlled current transformers — magnetic amplifiers, driven one way, the other way, or not at all. Three drive conditions, and they are real.
What is not true, and this page said otherwise until 5 September 2026, is that the core itself had three natural states. A square-loop ferrite core is deliberately bistable: it holds positive remanence or negative remanence, and zero flux is the unstable point between them rather than a third resting place. Setun's cores were not even ideally square, so compensation was needed — and the consequence is recorded in the machine itself. Each trit was stored in a pair of cores, wired in tandem so that the pair had three stable states. A trit took two tracks on the drum and two holes in the paper tape, for the same reason.
So Setun was exactly what this site says elsewhere it was: a three-valued code built on two-state parts, and built extremely well. Brian Hayes puts the sharper version of it — a pair of cores could have carried two bits, which is four states, against the three a trit uses, so at the storage layer the ternary advantage was not merely unassisted by the hardware, it was paid for.
This correction makes the argument of this post stronger rather than weaker, which is why it is worth making rather than quietly editing. If Setun had owned a natively three-state device, its abandonment would be evidence that having one is not enough — and the interesting question would move to economics. It did not own one. Nobody has. That is the gap. Sources: the element description is the Russian Virtual Computer Museum's account of Brusentsov's work; the pair-of-cores detail and the four-versus-three comparison are Brian Hayes, Third Base, quoted in Wikipedia's Setun article.
Then the element base changed. The transistor displaced the magnetic amplifier over the following decade, everywhere, for reasons that had nothing to do with radix: speed, size, power, and the fact that a planar process could print millions of them. And a transistor is a switch. It is on or it is off. A third state has to be synthesised on top of it — by splitting a voltage range into three bands, by stacking devices with different thresholds, or by encoding on current rather than voltage.
Every one of those costs noise margin. A fixed supply rail divided three ways leaves less distance between adjacent states than the same rail divided two ways, and margin is what buys yield, temperature range and tolerance to variation. That is the tax, it is charged per gate, and it is the reason that multiple-valued logic has spent seventy years being a good idea in papers and a bad trade in fabs.
The institutional history is real too — the Soviet computing programme turned toward reproducing an existing Western architecture at the end of the 1960s, and a machine that shared no software with anything else was a poor fit for that decision. But the institutional argument only had force because the technical argument had already turned. If ternary hardware had still been the cheaper way to build a machine, it would have been much harder to stop.
What Has Materially Changed
Three things, none of which existed in any usable form when Setun was abandoned.
A device that is natively three-state — for the first time
A metallic carbon nanotube conducts with essentially the same physics in either direction. Current one way, no current, current the other way is not an encoding imposed on the device; it is what the device does. This is not a restoration of the relationship Setun had with its cores; Setun never had it. Setun's engineers spent two cores to buy three states, and this heading said “again” until 5 September 2026 on the strength of a claim about ferrite that does not survive checking — see the correction above. One device, three states, is the thing that has not existed before. The device patent, Thatte1, is about exactly this, and NEGF quantum transport simulation is what confirmed the three states are distinguishable at room temperature rather than merely conceivable.
Photonics, which sidesteps the voltage-band problem entirely
The noise-margin tax is a consequence of encoding three states in one voltage range. Encoding instead on the phase of an AC drive, gated optically, does not divide a range into three — it uses a signed carrier plus an absence, which are separated by their nature rather than by a threshold. Whether that fully escapes the tax is an engineering question with real remaining difficulties, and the three that matter most are stated plainly on the hard problems page. But it is a different question from the one that defeated electrical multiple-valued logic, and that is the point.
A demand side that did not exist
In 1970 nobody outside a research group needed three-valued arithmetic. Today, ternary weight quantisation is ordinary practice in machine learning — models are routinely reduced to weights in {−1, 0, +1} because it works, on hardware that then has to fake those three values in binary. A natively ternary accelerator is not a solution looking for a problem any more; that is what Thatte10 addresses.
Picking the Thread Up, From the Other End
There is an honest asymmetry worth stating. Setun had working hardware and almost no software: a machine that existed, programmed close to the metal, in a world with no ternary language, no ternary compiler and no ternary operating system to port.
This project is the mirror image. The compiler and the microkernel exist and run today — ManiT compiles to a 27-trit instruction set, thatteOS boots on the emulator, and both are public under AGPL-3.0 with archived DOIs. What does not exist yet is the hardware. The twelve patent specifications describe it, simulation supports it, and no wafer has been through a fab.
Stating that plainly is not modesty; it is the only way the comparison is useful. Setun proves the arithmetic works in hardware. The software stack proves the toolchain and the operating system are constructible. Neither proves the other, and the gap between them is precisely the device.
What Setun should be remembered for is not that it was a curiosity. It is that a balanced ternary computer, built from the best available parts of its decade, was ordinary enough to be sold and used — and that when the parts changed, the machine did not fail an argument. It simply lost its element base. Getting one back is a device problem, and it is the problem this lab is working on.