The Three Budgets Behind Every Quantum Control System

A transmon, the standard superconducting qubit design, holds a usable quantum state for something on the order of 100 microseconds on a well-built device. A single-qubit gate on that same device takes roughly 20 nanoseconds. Everything an engineer needs to understand about quantum control lives between those two numbers: thousands of operations can fit inside the coherence window, the window is the time before the state decays past usefulness, and every one of those operations has to be right the first time, because a quantum state cannot be copied and re-run.

The hardware that fills that window is the quantum control system. It is the classical layer that turns compiled instructions into shaped analogue signals, digitises the faint responses that carry measurement results, and closes feedback loops fast enough to act on what it just measured. Vendors describe these systems by their parts, and the parts are easy to list: waveform generators, digitisers, timing distribution, an FPGA or two. We find that engineers get further, faster, by describing them a different way. A control system is three budgets being spent at once. There is a timing budget, a noise budget, and a wiring budget. Every architectural argument in the field is an argument about which of the three to overspend and where to recover it.

Where the control system stops

Before the budgets, a boundary worth fixing, because the terms get used loosely.

A control system works in volts, gigahertz, and nanoseconds. Its obligations are physical: emit this waveform at this frequency with this phase, sample that response, decide within a microsecond. Almost all of them are built around field-programmable gate arrays, or FPGAs, because an FPGA gives deterministic latency that a general-purpose processor cannot promise.

A quantum operating system works in abstractions. It schedules jobs, allocates qubits, talks to compilers, and hides the specific hardware from the application. QuTech’s QNodeOS, published in Nature in 2025, is the clearest published example of that layer treated as its own research problem: platform-independent programs running across physically different quantum processors.

The classical analogy that holds up is the network interface card and the network operating system. The card moves electrical signals on the wire. The operating system routes, allocates, and presents one interface regardless of what card is fitted. Both are necessary and they solve different problems.

Several commercial systems straddle the line deliberately. Quantum Machines pairs its OPX controllers with QUA, a language that carries scheduling and real-time branching that looks like operating-system work. Zurich Instruments’ LabOne Q sits above its instruments in a similar position. That blurring is not sloppiness, it is a bet: as error correction matures, more of what we currently call system software has to run inside the real-time loop, and the boundary will move down.

For the rest of this article, we stay below the line. Volts, gigahertz, nanoseconds.

The timing budget

Timing is the budget engineers underestimate first, because the individual numbers look generous and the composition does not.

A single-qubit gate needs a microwave pulse in the 4 to 8 GHz range with a specified amplitude, duration, and phase. That much a signal generator handles. The difficulty is relative timing. Two-qubit gates depend on the phase relationship between pulses arriving at different qubits, so the jitter between channels, meaning the run-to-run variation in when a pulse actually arrives, has to be held well below a nanosecond across a chassis that may carry hundreds of outputs. Jitter that would be invisible in a communications system shows up directly as reduced gate fidelity here.

Then the loop closes. Mid-circuit measurement, where a qubit is read partway through a program and the result decides what happens next, means the control system must digitise a readout signal, discriminate the state, and emit a conditional pulse while the rest of the register is still coherent. Error correction turns that from an occasional operation into the steady state.

Working the loop arithmetic

It helps to spend the budget on paper once.

Take a superconducting device with a 100 microsecond coherence time. A surface-code error-correction round needs a syndrome measurement, which is a parity check on a group of physical qubits that reveals where an error occurred without revealing the encoded information. Suppose the readout pulse and its integration take 500 nanoseconds. Cable transit and analogue front-end delay add a few hundred more. Decoding the syndrome, which means turning a pattern of parity flips into a statement about which correction to apply, has to happen next, and then the correction pulse has to be generated and delivered.

If the whole round completes in one microsecond, the device gets roughly a hundred rounds inside a single coherence time. If it completes in ten microseconds, it gets ten. The physics did not change. The classical engineering did, and it moved the machine between two entirely different regimes of what error correction can accomplish.

This is why decoders migrate into FPGAs and, increasingly, into application-specific integrated circuits, or ASICs. Software decoding on a host computer introduces a round trip and a scheduler, and neither offers a latency guarantee. At scale the problem gets worse rather than better, because syndrome volume grows with code distance and with qubit count simultaneously. A fault-tolerant machine performs syndrome decoding millions of times per second, continuously, for the entire duration of a computation.

The noise budget

The noise budget is where classical test-and-measurement instincts transfer most directly, and where they need one adjustment.

The relevant specifications are familiar. Phase noise, the short-term random fluctuation in an oscillator’s cycle timing, converts into accumulated phase error across a circuit. Spurious-free dynamic range, or SFDR, describes how far the largest unwanted spectral component sits below the intended signal; when SFDR is insufficient, a spur can land on a neighbouring qubit’s transition frequency and drive an operation nobody asked for. Amplitude flatness across the band determines whether a calibration performed at one frequency still holds at another.

The adjustment is this: in most RF work, the device under test is the noisy part and the instrument is assumed clean. In quantum control, that assumption expires. Coherence times have improved by orders of magnitude over the past decade, and on the best current devices the control electronics are a meaningful contributor to the total error. Zurich Instruments built its product line around that observation, using a double-superheterodyne conversion scheme that avoids the IQ mixer calibration drift common in simpler upconversion chains. Whether that specific approach wins is a market question. The underlying trend is not in dispute: as qubits get quieter, the instrument stops being a bystander.

Readout sits inside the same budget and deserves separate attention. A superconducting qubit is measured indirectly, by probing a coupled microwave resonator whose frequency shifts depending on the qubit state. The returning signal is extremely weak and normally passes through a chain of quantum-limited amplifiers, often Josephson parametric amplifiers, before it reaches the digitiser. The control system then applies matched filtering or weighted integration and produces one bit.

That bit is the whole point, and the consequences of getting it wrong are asymmetric. In a bare circuit, a misread produces a wrong answer that averaging can wash out over many shots. Inside error correction, a misread syndrome does something worse: it tells the decoder to apply a correction where none was needed, which injects an error into an otherwise healthy logical qubit. Readout fidelity therefore sets a floor underneath everything above it, and no amount of algorithmic work recovers what the discriminator threw away.

The wiring budget

The third budget is the one with the hardest wall.

Superconducting and silicon-spin qubits live inside a dilution refrigerator, a cryostat that holds the processor near 10 millikelvin. The control electronics sit at room temperature. Connecting the two requires coaxial cable, typically two or three lines per qubit for control plus one for readout, threaded down through the temperature stages of the fridge.

Do the arithmetic. A thousand-qubit machine needs several thousand cables. Each one is a thermal path from a warm environment into a cold one, and the cooling power available at the coldest stage is measured in microwatts. Each is also an attenuator, a delay line, and a candidate for crosstalk. Cable count is not an inconvenience at this scale, it is the scaling limit, and adding a zero to the qubit count is not something the current approach survives.

Three lines of attack are in progress, and they are not mutually exclusive.

The first is multiplexing. Readout already does this: many resonators are placed at different frequencies on one line, and one digitiser recovers all of them by frequency-division. Extending comparable sharing to control lines is harder, because control signals are not passive responses, but the same logic applies.

The second is moving the electronics into the cold. Cryo-CMOS, meaning complementary metal-oxide-semiconductor circuits designed to run at cryogenic temperatures, puts control functions on a chip operating a few kelvin above the qubits rather than three metres and 300 kelvin away. Intel’s Horse Ridge controllers were an early public demonstration of the idea at around 4 kelvin. A parallel line of work uses single-flux-quantum logic, a superconducting digital family, integrated with the qubit chip through flip-chip bonding, so the control pulses are generated millimetres from their target. Both approaches trade a new problem for the old one: every milliwatt dissipated at low temperature is a milliwatt the fridge has to remove.

The third is replacing copper with light. Optical fibre carries far less heat than coaxial cable per channel, and photonic interconnects at the fridge boundary are an active area for exactly that reason.

How modality shifts the balance

All three budgets exist for every qubit technology. Which one binds first is what changes, and it changes enough that control hardware does not transfer casually between modalities.

Modality	Primary control signal	Coherence, typical order	Budget that binds first
Superconducting	Microwave, 4–8 GHz, plus flux bias	~100 µs	Timing and wiring
Trapped ion	Laser, plus microwave	Seconds and longer	Sequence complexity, optical stability
Neutral atom	Laser, with spatial addressing	Hundreds of ms to seconds	Spatial control and imaging
Photonic	Optical routing and phase	Loss-limited rather than decoherence-limited	Phase stability and feed-forward speed
Silicon spin	Microwave, plus DC gate voltages	Up to milliseconds, isotopically enriched	Charge noise and calibration drift

Trapped-ion work has the longest coherence and therefore the loosest timing budget, and it spends the slack on sequence complexity instead. The community’s standard tool reflects that history. ARTIQ, developed by M-Labs with the Ion Storage Group at NIST, is an open-source control framework that compiles Python experiment code onto FPGA hardware with nanosecond timing resolution and sub-microsecond conditional branching. It runs on Sinara, a modular open hardware family released under the CERN Open Hardware Licence. An engineer joining a trapped-ion group is far more likely to be handed an ARTIQ crate than a commercial chassis, and the openness is a genuine advantage during hardware bring-up, when the ability to modify the control layer matters more than polish.

Neutral-atom systems move the difficulty into space. Atoms are held in optical tweezers, which are tightly focused laser beams, and arranged by spatial light modulators and acousto-optic deflectors into arrays. Gates are driven by exciting atoms into Rydberg states, highly excited states that make neighbouring atoms interact strongly. Readout is a camera image of fluorescence. The control problem is therefore geometric and optical before it is electronic, and Pasqal’s open-source Pulser framework reflects that: a program specifies an atom layout first, then pulse envelopes over time, rather than a gate list.

Photonic machines change the question again. Photons do not decohere in the way matter qubits do, they get lost, so the budget is loss and phase stability rather than coherence time. Control means keeping interferometers phase-locked, synchronising sources with detectors, and executing feed-forward, meaning a measurement result that changes a later optical operation inside the same run. Xanadu’s Aurora demonstrated that this can be automated well enough to run unattended for extended periods at room temperature, which is a control-engineering result as much as a physics one. PsiQuantum’s Omega chipset, manufactured on a commercial silicon photonics line, takes the co-location advantage further, since photonic components tolerate proximity to warm electronics in a way superconducting circuits do not.

Silicon spin qubits sit closest to the superconducting case and borrow its instruments, which is why groups such as Diraq at UNSW Sydney run commercial platforms. Their long-term problem is different, though. The reason to build qubits in a CMOS process is the prospect of millions of them on a die, and no cable-based control scheme reaches that number. Silicon spin is the modality where cryo-CMOS is not an optimisation but a precondition.

Buying back budget with pulse shaping

Once the three budgets are visible, an obvious question follows: which of them can be recovered in firmware rather than hardware?

A fair amount, as it turns out. A square pulse that rotates a qubit is the simplest thing to emit and among the most fragile, because it is sensitive to amplitude error, frequency drift, and crosstalk. A shaped pulse, designed using optimal control theory, can perform the same rotation while remaining far less sensitive to all three. The mathematics descends from decades of nuclear magnetic resonance work, and the modern difficulty is computational rather than conceptual: optimising pulses for a multi-qubit system under real hardware constraints is expensive, which is why Q-CTRL built a business around doing it in the cloud and shipping the results down to the control hardware.

The practical consequence for anyone specifying a system is that raw analogue quality and pulse intelligence are partial substitutes. A cleaner instrument reduces the error a pulse has to tolerate. Better pulse design reduces the cleanliness the instrument has to deliver. Neither eliminates the other, and the interesting engineering decisions live in the trade.

Reading a control system specification

Datasheets in this field are dense and selectively silent. A few questions cut through most of them, and we build the same habits with learners working through control hardware:

• What is the channel-to-channel timing skew and jitter, not the sample rate? Sample rate is the number vendors lead with. Relative timing across channels is what two-qubit gate fidelity depends on.
• What is the closed-loop latency from digitiser input to conditional output? Ask for the figure including the decision, not just the acquisition. Then compare it against the coherence time of the target device.
• How many readout channels can share one digitiser, and at what state-discrimination fidelity? Multiplexing ratios quoted without a fidelity figure are marketing.
• What runs on the FPGA, and can the customer change it? This determines whether a new error-correction scheme is a firmware update or a vendor request.
• How does the system behave as channel count grows? Timing distribution across multiple chassis is where architectures that looked fine at 20 qubits come apart at 200.

The last one generalises into a useful heuristic for evaluating a quantum computing platform as a whole. A company that talks in detail about its control architecture, whether built in-house or bought, has usually confronted the full engineering problem. A company that presents qubit counts and skips the control layer may be further from a scalable machine than the headline suggests.

Where to build this

Control engineering is where quantum computing stops being a physics story and becomes a systems-integration problem: mixed-signal design, deterministic real-time computing, thermal engineering, and calibration software, all constrained by physics that punishes imprecision immediately. Engineers arriving from RF, embedded, or test-and-measurement backgrounds already hold most of the required foundation. What they usually lack is the mapping between what they know and what the qubits demand, and that mapping is learnable in weeks rather than years.

Quantum Academy’s quantum engineering programs cover the control layer as an engineering discipline in its own right, from signal chain design through real-time feedback and calibration. For deeper technical background on specific vendor architectures and the research behind cryogenic control, PostQuantum.com’s treatment of quantum control systems goes further into the hardware detail than we do here. To see how the discipline fits into the broader set of quantum engineering roles, QuantumCareers.com maps the paths.

Start with the programs at quantumacademy.com/.