Hardware · Quantum Computing
IBM tests 80 cm quantum interconnects between two 120-qubit processors
IBM tests quantum interconnects between two Nighthawk processors. State transfer reaches 99%, but gates between chips still fail about 5% of the time.
Key points
- IBM connects two 120-qubit Nighthawk processors with four superconducting cables about 80 cm long, assembled with the modules already in the cryostat.
- State transfer between chips reaches 99.00% average fidelity; CZ gates between qubits on different chips, 94.75%.
- It is an unreviewed preprint with partially wired test chips: there is no error correction between modules.
A team at IBM Research has connected two 120-qubit superconducting processors with four quantum interconnects made of about 80 centimetres of cable, and has run logic gates between qubits on different chips. It describes the work in a preprint published on 21 October, not yet peer reviewed. According to the authors, it is the first multi-channel interconnect between many-qubit superconducting processors.
It is one building block of a modular architecture: instead of an ever larger chip, several modules joined by links that let them work as a single machine. Within each module, a dense network of qubits; between modules, fewer connections. According to the preprint, splitting the machine this way helps even if the links between modules have error rates an order of magnitude higher than the gates within the chip.
What they have built
- Two test Nighthawk R1 processors. To move faster, only the qubits around the links were wired up and measured, and fabrication steps that improve qubit coherence were left out.
- Four links. Each one joins a qubit on each chip through two dedicated tunable qubits, two couplers and a superconducting aluminium coaxial cable about 80 cm long. The cable meets the chip perpendicularly, so that more connections fit, and is attached to it with an indium cold bond made with both modules already installed in the cryostat.
- A new problem with the magnetic field. If the cable is cooled in the presence of a magnetic field, it loses quality: in the tests, the loss increased sixfold between zero field and 14 gauss, and it does not recover afterwards. For now it has been solved with two layers of magnetic shielding, which the authors regard as a temporary fix.
What results it gives
| Operation between the two chips | Average fidelity | Best link |
|---|---|---|
| State transfer | 99.00% | 99.29% |
| CZ gate between qubits on different chips | 94.75% | 96.42% |
With those gates they generate entangled states spread across the two processors: a Bell pair with a fidelity of 85.9% without correcting readout errors, and two at once over neighbouring links, which shows that the links work independently. Also four- and six-qubit GHZ states, at 77.2% and 67.7% without readout correction. All four links kept working after three cooling and warming cycles without touching the connections, something a machine that will need repairing many times requires.
Note · what this work does not show
This is not a working modular computer: these are two partially wired test chips with four links. Gates between chips fail roughly one time in twenty, and there is no error correction or protected logical operation between modules. The magnetic shielding is temporary, and the indium bond has not been taken apart and reassembled.
What changes for decision-makers
Scaling up superconducting quantum computers will require joining many modules with many nearly identical connections, and this work shows one way of building them at metre scale, on large processors and with the system already installed. According to the authors, most of the state-transfer error is explained by dephasing of the link qubits, whose coherence can be improved. For those following quantum computing timelines, including that of a computer able to break today’s cryptography, this is an engineering advance worth watching, not a change of schedule: these connections still have to be shown in complete machines and with much lower errors.