First-hand information for everyone
Main result Researchers demonstrated four long‑distance couplers ("L‑couplers") between two 120‑qubit IBM Nighthawk R1 superconducting processors. Each L‑coupler joins a pair of lattice qubits across the modules and uses an ≈80 cm superconducting coaxial link. (D1, M1)
Link construction Each L‑coupler comprises two frequency‑tunable L‑qubits, two RF‑SQUID–derived “i2m” tunable couplers, on‑chip transitions, and ≈80 cm of superconducting aluminum coax. The cables are installed perpendicular to the processor surface. Cable‑to‑chip joins were made after module mounting using an indium‑to‑indium room‑temperature pressure bond. The i2m coupler’s always‑on negative geometric mutual inductance produces a flux‑insensitive off point. (M1, M2, M3)
How performance was measured A single standing cable mode near 5 GHz (free spectral range ≈150 MHz) was isolated for relay‑based state transfer. Coupling and mode parameters give g/FSR ≈ 0.05–0.06. Transfer durations per link were 256–312 ns. State transfer fidelity was measured with network benchmarking. Each link was benchmarked in both directions; reported link fidelities are the averages. (M4, C1, M3)
Measured performance The four assembled cable modes near 5 GHz had internal quality factors Q_i = 0.95×10^6, 1.09×10^6, 0.86×10^6, and 0.86×10^6, with a mean of (0.94 ± 0.09)×10^6. (D2) Relay‑based state transfer across the four L‑couplers has a mean network benchmarking fidelity of 99.00%. All four links transfer at or above 98.86%; the best link is 99.29 ± 0.07%. (D3) The authors implemented distant controlled‑Z (CZ) gates between lattice qubits on separate modules. The distant CZ gates reach a mean fidelity of 94.75% with the best link at 96.42 ± 0.28%. The compiled distant CZ sequence follows prior work and combines three local CZs, two state transfers, four single‑qubit gates and one virtual phase correction; randomized benchmarking treats the compiled sequence as a single two‑qubit gate. (D4, C2) Using the distant CZ gates the team generated entangled states across modules. Examples: Bell pair BP 115–233 unmitigated F = 85.91 ± 0.40% (readout‑mitigated 93.94 ± 0.27%); BP 117–231 unmitigated 76.93 ± 0.49% (mitigated 86.98 ± 0.41%). Four‑ and six‑qubit GHZ states reached unmitigated fidelities 77.23 ± 0.67% and 67.67 ± 0.66% (mitigated 91.70 ± 0.70% and 90.87 ± 0.74%). (D5)
Significance and supported caveats A coherence‑only error model that includes only L‑qubit and cable‑mode coherences accounts for about 77% of the measured state transfer error. Using that model with L‑qubit T1 = T*2 = 100 μs bounds transfer fidelities to 99.68–99.77% on these links, indicating room for improvement via better qubit coherence and faster pulses. (C3) The authors state they intentionally skipped fabrication and processing steps known to improve qubit coherence to speed development. Thus the reported limited qubit coherence is not presented as a fundamental limit of L‑coupling and can be improved. (L1) Cable loss depends strongly on the magnetic field present while the aluminum cable cools through its superconducting transition. Measured median Q_i (4.5–7.5 GHz) versus cooldown field was 1.5×10^6 at 0 G, 1.1×10^6 at 1.4 G, 7.4×10^5 at 3.5 G, and 2.5×10^5 at 14 G. Changing the field after cooldown had no effect; the mechanism behind the loss is not yet determined. (L2) Magnetic shielding with a braid mitigated cooldown‑field loss in single‑cable tests (a shielded test cable retained median Q_i ≈ 2.0×10^6 near 5 GHz). In the assembled dual‑module setup all four cables shared a single braid; the four cable modes then reached the Q_i values above. The authors identify bundling and reduced shielding effectiveness as a leading candidate for the lower Q_i in that assembly. (L3, D2) For a representative distant CZ (CZ 111–237) a summed constituent error budget upper‑bounds the gate at 97.34 ± 0.05%. That budget is dominated by the two state transfers (≈53% of the budget) and the three local CZ gates (≈40%). The authors attribute much of the remaining discrepancy with the measured fidelity to residual population left on the L‑qubits by incomplete state transfer. (L4) Readout error is the dominant limitation on every reported entangled‑state fidelity in this setup. Readout mitigation raises the reported fidelities. The authors note the readout quality reflects this test system’s measurement chain rather than an intrinsic property of the L‑couplers or qubits. (L5)
Bottom line Four practical, multi‑channel L‑couplers were assembled between two 120‑qubit Nighthawk processors. They demonstrate high‑quality relay transfer (mean 99.00%, best 99.29 ± 0.07%) and distant CZ gates (mean 94.75%, best 96.42 ± 0.28%). The paper gives concrete hardware details, measured fidelities, and quantified limitations that point to clear routes for improvement, especially via improved qubit coherence, cooldown magnetic control, and shielding. (D1, D2, D3, D4, D5, M1, M2, M3, M4, C1, C2, C3, L1, L2, L3, L4, L5)