IBM plans to wire two of its next quantum chips together in 2027 and run them as one machine with 1,040 qubits, the company said on November 11. Each of the new processors, called Nighthawk r3, is designed for 520 programmable qubits and for circuits of 10,000 quantum operations, and each carries 10 chip-to-chip links that IBM calls l-couplers.
“Useful quantum computing hinges on whether results can be delivered with accuracy and speed, both are fundamental,” said Jay Gambetta, director of IBM Research, in the company’s announcement from Yorktown Heights, New York.
Why link two chips
A qubit is the quantum version of a bit, able to hold a blend of 0 and 1 until it is measured. Building one ever-larger chip gets harder as qubit counts climb, so IBM is betting on modules: smaller processors joined by couplers, sitting in cryogenic units that can themselves be connected.
IBM says both halves of that plan were tested this year. Earlier Nighthawk r1 chips, joined by four l-couplers, passed quantum states across with 99.3 percent fidelity; the goal is 99.9 percent. It also coupled a pair of cryogenic modules for the first time; the combined unit got cold enough to work in less than five days and dropped below 15 millikelvin.
Those numbers are easier to read with a little arithmetic. A 99.3 percent transfer rate means roughly 7 in every 1,000 hand-offs between chips go wrong; hitting 99.9 percent would cut that to about 1 in 1,000. And 15 millikelvin is about 180 times colder than the 2.7-kelvin glow of deep space, since 2.7 divided by 0.015 is 180.
- Nighthawk r2 (available since September): 120 qubits, more than 7,500 operations, up to 100,000 samples a second.
- Nighthawk r3 (next): 520 programmable qubits out of 2,034 physical qubits, 10,000 operations, 10 l-couplers.
- 2027 system: two r3 chips linked, 1,040 qubits.
What today’s chip is doing
Users of Nighthawk r2, among them the energy company E.ON and the software firms Algorithmiq and Q-CTRL, have pushed 34 billion circuits through it, IBM says, with reported gains as high as 19-fold versus IBM’s previous chips. Speed matters because a quantum computer answers in samples, one result per circuit run, and serious workloads need millions or billions of them. The r2’s fast qubit reset shortens the wait between runs.
In IBM’s examples, E.ON’s machine-learning work on power-grid optimization ran 15 times faster, and an Algorithmiq model of a mixed quantum material finished 8 times quicker than on the older Heron chip. IBM estimates the newer circuits would need about 240 times more memory to match with leading classical tensor-network methods.
Quantum plus classical
The company is also folding high-performance classical computing into its IBM Quantum Platform, so users can pass work between a quantum processor and anything from a laptop to a supercomputer. By IBM’s account, RIKEN researchers repeated an earlier workflow roughly 5 times faster with 100 times as many calculations, and Cleveland Clinic and RIKEN pushed 608 million samples through 33 times faster once the steps were automated.
The push follows Japan’s own step in the same direction, when the University of Osaka linked its Japan-made quantum computers to a campus supercomputer. Extreme cold is the common thread in this field; a Chinese team recently chilled atoms to 9.7 picokelvin aboard the Tiangong space station.
IBM has not given a month for the two-chip machine. Its stated target is 2027, with 520 qubits on each side of the link.
Further reading
Image: Irekia / Eusko Jaurlaritza, CC BY 4.0 via Wikimedia Commons
This article was produced with AI assistance and reviewed by our Editorial Desk. How we work.
