Making It Compute

Real Machines, Real Mess

9 min

Everything so far assumed perfect qubits that sit still and do as they are told. Real ones do neither.

Why circuits have to be short
0timecoherenceusable circuit depthresults wash out into noise
A qubit forgets. T₁ is how long it holds an excited state; T₂ is how long it holds a phase relationship. On today's superconducting hardware both sit in the hundreds of microseconds, and every gate spends some of that budget.

Qubits forget

Left alone, a qubit slowly loses whatever you put into it. On today's best chips it holds on for a few hundred millionths of a second.

Every gate you apply spends a slice of that. Run out, and the delicate pattern of signs you were building is gone — and with it, the answer.

1 in ~1,000

Roughly how often a two-qubit operation goes wrong today

1 in 100

How good it must be before error-fixing starts to help

A few hundred µs

How long a qubit remembers. Every gate spends some of it.

Nobody has won yet

Superconducting chips (IBM, Google) are tiny circuits chilled to colder than deep space. Fastest gates, shortest memory, and each qubit can only talk to its neighbours.

Trapped ions (Quantinuum, IonQ) are single atoms held still by electric fields. Beautifully accurate, any qubit can talk to any other — but the gates are about a thousand times slower.

Neutral atoms (QuEra, Pasqal) are atoms held in place by laser beams. Newer, and unusually good at scaling to large numbers.

Nobody is winning. They are all trading the same limited budget in different ways.

Fixing errors is expensive

Ordinary computers fix errors by keeping three copies and taking a vote. You cannot copy a qubit — that is a rule of the universe, not a missing feature.

The workaround spreads one good qubit across a whole grid of ordinary ones, and repeatedly asks questions like "do these two still agree?" — questions that reveal an error without revealing the answer.

It works. But only once your gates are already better than about 1-in-100, and it takes roughly a thousand physical qubits to make one reliable one.

Worth remembering

  • Qubits forget after a few hundred millionths of a second, and every gate spends some of that.
  • Superconducting, trapped-ion and neutral-atom machines all trade the same budget differently.
  • You cannot copy a qubit, so errors are fixed by spreading one across many.
  • That costs around a thousand ordinary qubits per reliable one.
  • Qubit counts in headlines say much less than they seem to.