A quantum research task that took 12 weeks now takes 45 minutes, Santa Clara’s Nvidia says, as it teams with IBM and Fermilab to learn how to link quantum chips together
The three plan a test bed that pairs Nvidia AI supercomputers with IBM quantum computers. Nvidia also says supercomputer centers can now book quantum processors as easily as regular chips.
Nvidia, the Santa Clara chipmaker, said Monday that new software cuts one of the slowest parts of quantum computing research from 12 weeks to 45 minutes. In the same announcement, Nvidia, IBM and the U.S. Department of Energy’s Fermi National Accelerator Laboratory near Chicago said they plan to build a shared test bed for figuring out how to link several quantum computers so they can work as one.
Both are research steps, not products people can use. The platform is still a plan: the partners say they “intend” to build it and haven’t said when it will open, where the machines will sit or what it will cost, according to Fermilab.
What it means
Quantum computers are expected someday to tackle problems that stump today’s computers, such as simulating molecules for new drugs and materials. But today’s machines make too many errors to do that reliably. These announcements target the two biggest hurdles: correcting errors and growing machines large enough by connecting smaller ones.
How it works
A regular computer stores information in bits that are either 0 or 1. A quantum computer uses qubits, which can hold a blend of both, but they are fragile and easily disturbed. Error correction spreads each piece of information across many qubits, and a regular computer program called a decoder reads measurements to figure out which errors happened so they can be fixed, Nvidia explains. Its new decoder, called TCBP-JOBS, is part of its free CUDA-Q software and speeds the heavy math of testing error-correction designs by up to 13,000 times, the company says.
What took 12 weeks, and why it now takes 45 minutes
The 12 weeks is a computer test, not a lab experiment. To choose an error-correction design for a quantum chip, researchers simulate the chip millions of times on ordinary computers with random errors mixed in, and check whether a decoder can work out what went wrong each time. Each simulated run is called a shot, and how often the decoder gets a shot wrong shows how good the design is, Nvidia’s research paper explains.
Nvidia’s benchmark was a test set that Google Quantum AI published for its own highly accurate decoder, called Tesseract: 24.2 million shots across 28 circuit designs. Tesseract needed 130,257 hours of processor time to get through them, which the paper says equals about 12 weeks on a 64-core server. Nvidia’s new decoder did the same set in 45 minutes on one Nvidia GB200 graphics chip, with about the same accuracy, the paper says.
The time savings come from smarter guessing. Tesseract searches for the single most likely set of errors, and that search grows explosively as codes get bigger, according to the paper. Nvidia’s method first runs a quick, rough pass, known as belief propagation, and learns from it how likely each possible fault is. It then uses those odds to steer many short searches at once on a graphics chip, which is built to run many calculations side by side. The speed and accuracy figures are Nvidia’s own, measured on its hardware.
The planned platform would put Nvidia DGX B200 systems, the company’s AI supercomputers, next to IBM quantum computers, including ones built on Nighthawk, a 120-qubit processor that can run up to 7,500 two-qubit operations, IBM said. Researchers would first simulate the links between quantum processors to learn how errors and delays in those links affect results, then use what they learn to design real connections. Fermilab would bring in physicists to try problems such as particle collisions, neutrinos and the insides of atomic nuclei.
Separately, Nvidia and IBM said they have added quantum processors to Slurm, free software that many of the world’s largest supercomputers use to decide which jobs run on which chips. That lets a supercomputer center book time on a quantum processor the same way it books regular processors and graphics chips, instead of running a separate system.
“Making quantum computers useful for science requires advances in error correction, applications and the systems that bring quantum processors and accelerated computing together,” said Krysta Svore, Nvidia’s vice president of applied research for quantum computing.
Source: Nvidia, announcement of collaboration with IBM and Fermilab and new decoding algorithm, Nov. 16, 2026; IBM, “IBM, Fermilab, and NVIDIA Announce Plans for New Co-Design Platform,” Nov. 16, 2026; Fermilab, “Fermilab, IBM and NVIDIA plan modular quantum computing co-design platform,” Nov. 16, 2026; Nvidia Research, “TCBP+JOBS: GPU-Accelerated Quantum Error Correction Decoding and Noise Learning with Near-Optimal Accuracy,” 2026.
- Nvidia Newsroom: Collaboration with IBM and Fermilab, TCBP-JOBS decoder (Nov. 16, 2026)
- IBM Newsroom: IBM, Fermilab and NVIDIA co-design platform (Nov. 16, 2026)
- Fermilab News: Fermilab, IBM and NVIDIA plan modular quantum computing co-design platform (Nov. 16, 2026)
- Nvidia Research, TCBP+JOBS: GPU-Accelerated Quantum Error Correction Decoding and Noise Learning with Near-Optimal Accuracy (paper, 2026)