Cornell's chip cools a trapped ion in 150 microseconds, and published time budgets show how much of a computation that step could reclaim
Published time budgets show cooling takes a quarter to two-thirds of some trapped-ion run times. Here is what a threefold faster step could and could not save.
In timing data Quantinuum has published for its trapped-ion processors, cooling the ions back down takes far longer than the logic gates. For the example circuits in its 2023 paper on the H2 machine, cooling took between 25 and 68 percent of run time and gates took 1 to 2 percent. So a faster cooling step could shorten whole computations, not just one stage. The saving is capped by cooling's share of the run, though, and it shrinks or disappears wherever cooling already runs at the same time as ion shuttling.
The faster method comes from Karan Mehta's group at Cornell's Duffield College of Engineering. It was published in PRX Quantum on November 9, 2026. Laser light delivered through waveguides built into the ion-trap chip brought a single trapped calcium ion close to its quantum ground state (its lowest possible energy of motion) in about 150 microseconds. A conventional laser arrangement took about 500 microseconds. The team used electromagnetically induced transparency (EIT) cooling, in which one laser makes the ion transparent to a second beam except at frequencies that remove vibrational energy. It cooled all three measured directions of motion, which span about 5 megahertz in frequency, in one step. Mehta's group says the ion ended up in its motional ground state about 95 percent of the time in the targeted direction.
Why ions keep needing to be cooled
Trapped-ion computers hold charged atoms in electromagnetic fields. The Cornell announcement explains that even tiny unwanted motion interferes with calculations, so lasers have to cool the ions near their lowest energy state again and again. The Cornell paper says high-fidelity laser-driven gates often need very low phonon numbers (a phonon is one quantum of vibration) in several modes of motion at once.
Motion builds up from several sources. Leading machines use the quantum charge-coupled device (QCCD) design, which moves ions between separate zones for storage, gates and measurement. The Cornell authors say these machines cool repeatedly to undo motion added by moving and rearranging ions, and to counter heating that happens unavoidably during operation. In Honeywell's 2021 QCCD demonstration, published in Nature, each transport step lasted 50 to 300 microseconds and added roughly as much motion as Doppler cooling leaves behind. (Doppler cooling is the first, coarser stage of laser cooling.) Measurement also heats ions. In the Cornell experiment, readout left about 100 phonons in one mode before cooling began.
Commercial machines mostly cool sympathetically: they cool a second ion species held in the same crystal, which shares its motion with the qubit ions, so the qubits themselves are not disturbed. Honeywell's demonstration paired ytterbium-171 qubit ions with barium-138 coolant ions. Quantinuum's Helios reverses the arrangement and uses ytterbium-171 to cool barium-137 qubits.
Where the time goes
Honeywell's 2021 paper gives a time budget for a teleported CNOT, a two-qubit operation carried out across ions in different zones. The authors called these cooling protocols the device's current runtime bottlenecks.
The H2 paper breaks down 11 example circuits on up to 32 physical qubits. Cooling took 47 percent of a 16-qubit quantum volume test, 41 percent of a repetition-code error-correction run and 68 percent of a two-qubit benchmarking run. Ion transport took most of the rest, about 60 percent on average.
Quantinuum's preprint on Helios, a 98-physical-qubit machine, describes a design change aimed at this bottleneck. In H1 and H2, cooling and gates shared the same zones, and the company says cooling was up to two orders of magnitude slower than gating there. In Helios, ions are cooled in the operation zones, by sideband sequences of about 3 milliseconds, while the next batch of qubits is sorted in parallel. The paper reports an average of 55 milliseconds per layer of a random 98-qubit program. Quantinuum says it will focus next on cutting transport time.
What a threefold speedup would and would not buy
A simple calculation gives an upper limit. Cutting cooling from 500 to 150 microseconds leaves it at 30 percent of its former length. If cooling is a share f of a sequential run, total time falls by 0.7 times f. Applied to the H2 figures, the two-qubit benchmark would be about 48 percent shorter, the repetition-code run about 29 percent shorter, and a QAOA optimization circuit that reuses qubits (cooling share 25 percent) about 18 percent shorter. These numbers are this article's illustration, not a claim made by Cornell or Quantinuum.
The real gain would likely be smaller, for four reasons. First, Cornell's 500-microsecond baseline was the same EIT method delivered as an ordinary travelling beam, not the sideband cooling Quantinuum uses. Second, the experiment cooled one ion in a trap held at 4.8 kelvin. The authors describe how the method could extend to multi-ion crystals containing coolant ions but have not shown it. Third, the authors list integrating the remaining pump beam onto the chip as future work. Fourth, in a design like Helios, where cooling overlaps with sorting, faster cooling shortens a layer only if cooling, rather than transport, is the step everything else waits on.
There is a second possible payoff. The Cornell authors say that reaching very low phonon numbers in a single cooling step could reduce cooling overhead and, at the occupancies the method allows in principle, could eliminate the need for sideband cooling. They also note that microwave-driven gates relax cooling requirements, so fast cooling affects how designers weigh laser gates against microwave gates.
The clock cycle and the decoder
Quantum error correction repeatedly measures syndromes, parity checks that flag errors without reading the stored data. A classical program called a real-time decoder must turn those syndromes into corrections faster than they arrive. The authors put error-correction cycles for trapped ions and neutral atoms at a few to tens of milliseconds. They compare this with about 1 microsecond for superconducting qubits.
A fast, less accurate decoder handles each window of syndrome data and hands off fewer than 1 window in 100 to a slower, more accurate one. With noise models based on IonQ's published blueprint and a physical error rate of 0.2 percent, the authors report logical error rates up to four orders of magnitude lower than with the fast decoder alone, at latency comparable to an existing decoder called Beamsearch.
This is where faster cooling meets decoding. A shorter cycle also leaves the slow decoder less time to work. The decoder work does not model cooling, so the public record does not say how much of its accuracy gain would survive a faster cycle. The authors do report that as physical error rates fall, the slow decoder is needed less often. Two things would settle the question: processor-level time budgets measured with integrated cooling on multi-ion crystals, and decoder simulations run at the cycle times those budgets produce.
- Researchers develop faster laser cooling for quantum computers, Cornell Chronicle
- Rapid multi-mode trapped-ion laser cooling in a phase-stable standing wave (arXiv v4 of the PRX Quantum paper), arXiv / PRX Quantum
- Preprint on experimental trapped-ion laser cooling, Cornell Photonics and Quantum Electronics Group (Mehta lab)
- Probing the real-time decoding accuracy bounds in trapped ion architectures, arXiv (preprint, not yet peer reviewed)
- Demonstration of the trapped-ion quantum CCD computer architecture (arXiv version of Nature 592, 209), arXiv / Nature
- A Race Track Trapped-Ion Quantum Processor (arXiv version of Phys. Rev. X 13, 041052), arXiv / Physical Review X
- Helios: A 98-qubit trapped-ion quantum computer, arXiv (Quantinuum) (preprint, not yet peer reviewed)