How Did Neutral-Atom Quantum Computers Become Capable of Executing More Than 100 Circuits per Second?
Quantum computing progress is often measured by qubit count, gate fidelity, or coherence time. But another metric is becoming increasingly important: how quickly a quantum processor can repeat a complete experiment. A system may contain high-quality qubits and perform accurate quantum gates, yet still have limited practical throughput if preparing, measuring, and resetting those qubits takes too long. A new experiment with a 10-qubit rubidium-87 neutral-atom array demonstrates a different approach. By combining nondestructive readout, atom reuse, a three-dimensional photonic-chip interface, and real-time FPGA processing, researchers achieved a raw quantum circuit iteration rate (qCIR) of 101 Hz, with a post-selected rate of 74.8 Hz. More importantly, the system achieved a normalized Fisher information rate of 57.7 Hz, more than an order of magnitude above the corresponding conventional approach.
Key Takeaways
* The bottleneck was not only the quantum gates. Neutral-atom processors can have excellent coherence and gate fidelity, but repeatedly preparing and measuring atomic qubits can severely limit how many complete experiments can be performed per second.
* Nondestructive readout allows the same atoms to be reused. Instead of losing the atoms during measurement and rebuilding the array, the researchers designed a readout process that preserves the atoms well enough to perform repeated circuit cycles.
* The experiment reached 101 Hz raw qCIR. Each repeated cycle takes approximately 7 milliseconds, allowing the processor to execute more than 100 circuit iterations per second under the demonstrated benchmark conditions.
* The optimal measurement point was not the highest-fidelity measurement. At a 1.5 ms readout time, the system achieved a median readout fidelity of 93.9% while retaining atoms with a median probability of 99.7%.
* Post-selection reduces the usable rate to 74.8 Hz. The researchers retained the 100-cycle datasets only when the initial and final atom-existence checks confirmed that the required atoms remained present.
* Information throughput improved even more dramatically than raw circuit speed. The normalized Fisher information rate reached 57.7 Hz, compared with only 3.37 Hz for the same system without atom reuse, even assuming perfect readout fidelity.
* Faster repetition accelerates quantum processor characterization. Randomized benchmarking produced an average single-qubit gate fidelity of 99.93%, with each data point acquired in approximately 13.2 minutes.
The Context
Neutral-atom quantum computing has become an important approach to building scalable quantum processors. Individual atoms can be trapped using optical tweezers and arranged into programmable arrays. Recent work has demonstrated increasingly large atom arrays and high-fidelity quantum gates, making neutral atoms promising candidates for future fault-tolerant quantum computing.
However, the number of qubits and the fidelity of individual operations do not tell the whole story.
A quantum experiment usually requires many repetitions. A circuit is prepared, the qubits are manipulated, the final state is measured, and the process is repeated many times so that probabilities, expectation values, or other statistical quantities can be estimated.
This makes the quantum circuit iteration rate, or qCIR, an important system-level metric. It represents the number of complete circuits executed per second, from initialization through quantum operations and readout.
For neutral-atom processors, one major obstacle has traditionally been the preparation of the atom array. Individual atoms are loaded into optical tweezers stochastically because of collisional blockade. Creating a defect-free array can therefore require loading, imaging, and rearranging atoms, with the preparation process taking hundreds of milliseconds.
Readout introduces another bottleneck.
Neutral atoms are commonly measured through fluorescence. The atom scatters photons that contain information about its internal quantum state. However, each scattered photon also transfers momentum to the atom. Collecting more photons generally improves state discrimination, but the accompanying recoil can heat the atom and increase the probability that it escapes the optical trap.
This creates a fundamental trade-off between measurement fidelity and atom survival.
A fast, destructive measurement can force researchers to reload and rearrange the atomic array. Alternatively, the measurement can be made less destructive by using longer detection and cooling processes, but that reduces the overall circuit iteration rate. Historically, these limitations kept qCIR below roughly 10 Hz in atom-array experiments, even when atom reuse was attempted.
The new experiment attacks this problem at the level of the complete experimental workflow rather than focusing exclusively on the quantum gates themselves.
The Main Idea
The central idea is straightforward: if the atoms survive measurement, there is no need to rebuild the quantum processor after every circuit.
The researchers use a one-dimensional array containing ten rubidium-87 atoms trapped in optical tweezers. The qubit states are encoded in two hyperfine states of the atomic ground-state manifold. Global optical and microwave fields are used for initialization, quantum operations, and readout.
The major technological component is a three-dimensional optical mapping chip, sometimes described in the paper as a “volcano” architecture.
The chip contains ten single-mode optical waveguides corresponding to the ten atom positions. At one end, the waveguides are arranged to match the closely spaced atom array. At the other end, they are separated sufficiently to connect to a fiber array and individual single-photon detectors.
This creates a direct optical pathway from each atomic qubit to its own detection channel.
The fluorescence collected from the atoms is therefore separated spatially, routed through the photonic chip, detected by independent single-photon detectors, converted into electronic signals, and processed in real time using an FPGA.
This architecture provides several advantages at once:
1. Parallel readout: multiple atom sites can be measured simultaneously.
2. Site-resolved detection: the signal from each atom can be associated with its corresponding qubit.
3. Fast electronic processing: the FPGA processes detection events in real time.
4. Atom reuse: the measurement is sufficiently nondestructive that the same atoms can participate in subsequent circuit cycles.
The researchers tested the approach using a sequence of 100 consecutive cycles. After the initial atom-loading and existence-detection stages, initialization, circuit execution, and readout were repeated 100 times before another existence check was performed. Each cycle lasted approximately 7 ms.
The Measurement Trade-Off
The key optimization is the choice of readout duration.
If the researchers increase the readout time, more photons are collected. This makes it easier to distinguish the two qubit states and therefore improves readout fidelity.
But the additional photon scattering also increases recoil heating. As the atoms become hotter, their probability of escaping the optical traps increases.
At a shorter readout time of 1.5 ms, the system sacrifices some measurement fidelity in exchange for much better atom survival. At this operating point, the median readout fidelity is 93.9%, while the median probability of retaining an atom is 99.7%.
This is an important conceptual point. The researchers are not optimizing for the highest possible measurement fidelity in isolation. They are optimizing the overall information throughput of the processor.
From 7 Milliseconds to 101 Hz
A 7 ms cycle corresponds to a theoretical repetition rate on the order of hundreds of cycles per second. Under the experimental benchmark sequence, the researchers obtain a raw qCIR of 101 Hz.
The raw rate, however, does not mean that every 100-cycle dataset can automatically be accepted.
The researchers therefore perform a conservative post-selection step. A sequence is retained only when both the initial and final atom-existence measurements confirm that the atoms remain present. After this filtering, the effective qCIR becomes 74.8 Hz.
The distinction between these numbers is important:
Raw qCIR: 101 Hz
Post-selected qCIR: 74.8 Hz
The first describes the demonstrated circuit-cycle speed, while the second accounts for the experiment's atom-presence selection criterion.
Why It Matters
The significance of the experiment goes beyond crossing a round-number threshold of 100 circuits per second.
Quantum computing experiments are inherently statistical. A single execution of a circuit rarely provides enough information to determine a result with high confidence. Researchers often need thousands or millions of samples.
If each sample requires a long preparation and measurement cycle, experiments become slow even when the underlying quantum operations are fast.
Atom reuse changes this equation.
Instead of spending a large amount of time rebuilding the atomic register after every measurement, the processor can keep using the same physical qubits for many consecutive experiments. The paper demonstrates 100-cycle sequences and shows that the repeated measurements remain statistically consistent.
The Rabi-oscillation measurements are particularly useful here. The oscillations obtained at the 25th, 50th, 75th, and 100th repetitions show consistent characteristics, including their bounds, frequency, and phase. Combining the repeated measurements substantially reduces statistical uncertainty.
The researchers further analyzed the inverse population variance as the number of repetitions increased. Its approximately linear growth indicates that the different repetitions contribute consistent information and can be combined without introducing additional bias.
Information Rate Can Matter More Than Circuit Rate
A particularly interesting aspect of the experiment is that the researchers do not evaluate performance using qCIR alone.
A circuit executed quickly is not necessarily useful if its measurements are too noisy.
To capture both speed and measurement quality, they use the normalized Fisher information rate:
Q = Rpos(2F − 1)²
where Rpos is the post-selected circuit iteration rate and F is the readout fidelity.
At the selected operating point:
F = 93.9%
Rpos = 74.8 Hz
Q = 57.7 Hz
This is compared with a value of only 3.37 Hz for the same system without the nondestructive-readout and atom-reuse strategy, even under the assumption of perfect readout fidelity.
The result therefore demonstrates something broader than faster measurement: a better experimental strategy can dramatically increase the amount of useful quantum information collected per unit time without replacing the quantum processor hardware.
Faster Benchmarking
Higher throughput also makes quantum processor characterization more practical.
The researchers performed single-qubit randomized benchmarking on the ten-atom array. The experiment used randomly selected gates from the set of ±π/2 rotations around the X and Y axes. Each data point used 593 random sequences averaged over ten atoms.
The resulting average single-qubit gate fidelity was 99.93%.
More importantly from a throughput perspective, each data point required approximately 13.2 minutes, demonstrating how faster circuit repetition can reduce the time required to characterize quantum operations.
This matters for quantum engineering because benchmarking is not a one-time task. Quantum processors must be repeatedly characterized while researchers optimize control pulses, hardware, calibration procedures, and error-management techniques.
What To Watch Next
The demonstrated system is still a relatively small processor containing ten atoms, so the result should not be interpreted as demonstrating a 100 Hz operating rate for a large-scale fault-tolerant quantum computer.
The experiment also shows that speed and fidelity remain linked. The 93.9% readout fidelity was deliberately chosen as part of an overall throughput optimization rather than as the maximum achievable measurement fidelity. Longer readout can improve state discrimination but increases the risk of atom loss.
Another limitation is the current cycle time. The researchers report that a complete iteration takes approximately 7 ms, with about 5 ms associated with cooling and 1.5 ms with readout.
The next major goal is to reduce these times dramatically.
The researchers suggest that improvements in photon-collection efficiency and advanced readout techniques that suppress recoil heating could potentially reduce both cooling and readout times toward approximately 0.1 ms. If achieved, this could provide a route toward kilohertz-scale qCIR, corresponding to thousands of circuit iterations per second.
The architecture could also be extended beyond the demonstrated one-dimensional ten-atom system. The researchers identify two-dimensional arrays, dynamic circuits, mid-circuit measurements, and real-time feedback as important future directions.
The broader question is therefore no longer simply:
How many qubits can a quantum computer control?
It is increasingly becoming:
How efficiently can those qubits be prepared, operated, measured, and reused to produce useful quantum information?
That shift toward throughput-oriented quantum computing could become increasingly important as quantum processors scale.
