Quantum computers promise to one day solve problems beyond the most powerful supercomputers imaginable. But it’s often underappreciated how much classical computing it takes just to operate these machines. As qubit counts rise, innovations in this supporting infrastructure will be essential if they’re to live up to their promise.

To prepare for the scale of quantum computers the industry is working toward, many companies are also gearing up the classical hardware, and software, required to support them. In April, Nvidia announced new AI-based software to accelerate the classical tasks that enable quantum computers. Sydney-based quantum software company Q-CTRL has developed an automatic calibration algorithm for quantum computers, and is now leveraging Nvidia’s agent-based system. Other companies, including IBM Quantum, Cambridge, England–based Riverlane, which develops quantum-error correction, and Google Quantum AI, are developing similar tools.

The Role of Classical in Quantum

Digital computer chips are marvels of engineering, operating flawlessly out of the box and capable of trillions of operations without error. The quantum bits, or qubits, at the heart of a quantum computer, by contrast, are temperamental and unreliable, requiring regular calibration and complex error-correcting schemes to keep them on track.

Calibration and error-correction are fundamentally classical, not quantum, problems, and they require dedicated classical hardware to solve. As quantum computers get bigger, the scale of those resources will need to rise in lockstep. That means that for the foreseeable future, quantum computers are going to be hybrid devices with a healthy dose of classical computing on the side.

“The cheapest and fastest way to execute most computer programs is to run them on a classical computer—even if a quantum computer is available,” says Adam Zalcman, a quantum software engineer at Google Quantum AI. “This is true of most of the information processing involved in running a quantum computer itself.... Therefore, I expect that every practical and efficient quantum-computer architecture will incorporate fast classical devices.”

Tuning Quantum Hardware

While the transistor has cemented its place as the foundational component of classical chips, the qubits at the heart of a quantum computer come in many flavors—superconducting circuits, trapped ions, neutral atoms, even individual photons. Using them for computation requires a painstaking calibration process to turn the “bare metal” of the underlying hardware into a qubit that can be controlled to run quantum circuits, says Jay Guilmart, lead product manager at Q-CTRL.

Calibration has two stages. The first, known as “bring up,” determines the frequency at which each qubit resonates, how long it holds its quantum state, its sensitivity to control pulses, and the strength of its interactions with neighboring qubits. All of these factors determine its error propensity and response to control signals.

Done by hand, the process still requires someone with a Ph.D. and can take days or even weeks, says Guilmart. This isn’t a scalable solution and so there’s a growing drive to automate the process. This is challenging because every step relies on results from the previous step. So rather than relying on a predefined script, Q-CTRL has therefore built intelligent calibration software that examines the result of each measurement, diagnoses failures, and adjusts the approach before retrying.

“After each step, we analyze that data and we say, are we okay to proceed to the next step? Do we have to go back to the previous step? Do we have to re-recreate this step?” says Guilmart.

Calibration is also not a one-and-done process: key parameters drift over time, gradually degrading performance. Q-CTRL’s software performs “runtime recalibration” to nudge things back into place, but there’s a limit to how much on-the-fly adjustment is practical.

“If I’m running a recalibration, I’m not running a circuit,” he says. “Even though I’m maintaining some high system state and high fidelities, if it takes all of my uptime it’s worthless.”

Decoding Errors in Real Time

Even a well-calibrated quantum computer remains fault-prone, which is why companies are investing heavily in quantum error correction (QEC). This typically involves encoding quantum information across large numbers of physical qubits in their shared state—a “logical qubit“—so that errors in individual qubits can be detected and compensated for without destroying the encoded information.

Because measuring a qubit directly collapses its quantum state, errors are detected via parity checks, which query whether pairs of qubits share the same state. This produces a series of measurements known as a “syndrome,” which classical algorithms called decoders analyze to locate errors.

The process must happen extremely quickly. While many errors can be logged and corrected mathematically after an operation, some must be fixed immediately before the algorithm can proceed. Superconducting and silicon spin qubits can hold their quantum states only for microseconds or milliseconds, so errors must be decoded and corrected within that window.

These tight requirements mean decoders typically run on specialized silicon like field-programmable gate arrays (FPGAs) or application-specific integrated circuits (ASICs) optimized for speed, says Jerry Chow, CTO of quantum-centric supercomputing at IBM. “You need to be able to keep up and you need to be able to effectively decode on the fly,” he says. “The best way to do that is through very tightly integrated FPGA or ASIC decoder capabilities.”

To AI or Not to AI

There is growing interest in using AI to simplify quantum hardware control. In April, Nvidia released two models targeting calibration and decoding. The first uses a vision-language model to analyze calibration-measurement outputs—typically plotted as graphs—and passes that evaluation to an AI agent that decides how to tweak the processor. The second uses a convolutional neural network to identify the simpler, localized errors that make up the bulk of faults. More complex errors are passed to a traditional algorithmic decoder, but the first pass reduces computational load enough to deliver a 2x speedup.

The attraction of AI for decoding, says Sam Stanwyck, director of quantum product at Nvidia, is that while models are time-consuming to train, they are extremely fast at inference—and thanks to parallelization across many chips, that speed holds even as qubit counts grow.

But offloading to a GPU still introduces significant latency, says Marco Ghibaudi, vice president of engineering at Riverlane. “You can have a really fat pipe, but it’s really long,” he says. “Our job [approach] has always been to try to remove as many unnecessary steps and shorten the pipe, and then make every section of the pipe as fast as possible.”

IBM’s Chow agrees that GPU latency currently makes them infeasible for real-time decoding. He’s also cautious about AI for calibration, given its computational expense. The approach holds promise for understanding the physics of novel architectures or new kinds of circuits. But for well-characterized devices where you’re simply looking for small deviations, simpler physics-informed techniques can be considerably cheaper.

The two approaches aren’t mutually exclusive, however, says Google’s Zalcman. Neural networks excel at discovering hidden patterns in syndrome data that help identify complex errors algorithmic decoders sometimes miss. Google is therefore developing a hardware architecture that can incorporate both traditional and AI-based decoders, including its AlphaQubit 2 model.

In the long run, Andi Gu, a Harvard Ph.D. student working on AI decoders, thinks “the bitter lesson” will come for decoding. This refers to AI pioneer Richard Sutton’s argument that general-purpose learning methods consistently outperform handcrafted algorithms over time. “If you make the model large enough and you throw enough training data at it, it will learn to capture the hidden correlations better than any other handwritten algorithm,” says Gu.

Latency remains a barrier, but his group is researching ways to make AI decoders more efficient and smaller so that they can fit on an FPGA, cutting response times. This can degrade accuracy though, so finding the right balance is still a work in progress.

Regardless of which approach wins out, one thing is certain—future quantum computers will require massive classical support. Decoding is a continuous, computationally expensive process whatever technique you use, says Gu, so you will need a “healthy chunk” of classical hardware dedicated to that task.

Calibration compute overheads will similarly “blow up” as devices scale to thousands or millions of qubits, says Q-CTRL’s Guilmart. Current techniques are unlikely to scale, he adds, so new approaches will be needed. “We’re going to have to rearchitect and do things differently when we get to even 1,000 qubits,” he says. “So no one’s winning the battle today.”