NVIDIA Strengthens Its Quantum Connection

Quantum computers are one high-profile technology that NVIDIA does not build—which is kind of a relief, considering the company reaches into every corner of AI. NVIDIA does build supercomputers, though, and that has set up the company’s role in the quantum computing world.

NVIDIA is providing a literal connection to quantum computers. A few weeks ago at the Washington, D.C., GTC event, NVIDIA announced NVQLink, a networking connection between a quantum computer and a supercomputer. And at SC25 this week, the company announced that roughly 20 supercomputer centers around the world are adopting NVQLink and/or aided in its development.

NVQLink is an open reference architecture, intended to work with any type of supercomputer and any type of physical qubits. The latter part is tricky because there are so many possibilities—qubits can be photons, neutral atoms, or ions, among other possibilities—so NVQLink’s development has involved 17 quantum computer partners so far, with more on the way, NVIDIA promises.

Error-Prone Qubits

Supercomputers and quantum computers will work hand-in-hand as the quantum side matures, but one immediate application for NVQLink is in quantum error correction (QEC).

Qubits are error-prone and sensitive to noise. So, to produce one logical qubit (a usable unit of information), a computer needs to generate dozens or even hundreds of physical qubits. Quantum error correction—handled by a supercomputer performing classical computing—applied to that collective of physical qubits produces one coherent logical qubit.

It’s part of the overall quantum race to produce machines large enough to be useful. That means scaling the number of physical qubits, but it also means a lot of attention is going to QEC. “It’s more accurate to say we’re building quantum error correction machines with a little bit of quantum computing,” said Sam Stanwyck, NVIDIA’s group product manager for quantum computing, in a recent briefing with analysts.

A Different Kind of Interconnection

NVIDIA’s initial product for connecting a quantum computer and a supercomputer was called DGX Quantum, built with the startup Quantum Machines as a partner. NVQLink is more generalized and open. (It’s also a direct upgrade from DGX Quantum, NVIDIA said.)

If you’re wondering what the industry did before DGX Quantum, NVIDIA notes that QEC at scale is a relatively recent pursuit. If all you’re doing is sending qubit results to a computer in a lab setup, Ethernet works well enough.

QEC, on the other hand, requires deterministic communication and reactions in the range of microseconds—hence, the need for specialized interconnection and the need for NVIDIA to enlist partners for the quantum side of the connection.

So, while NVQLink’s actual GPU-to-QPU throughput might seem modest—individual links run RDMA over Converged Ethernet (RoCE) at 400 Gbit/s—it handles real-time communications in ways that vanilla high-speed Ethernet can’t, such as providing GPU-to-QPU latency of less than 4 microseconds.

NVQLink’s mission goes beyond QEC. Calibration—the fine-tuning that’s required for qubit fidelity—is another example. Calibration resembles a series of experiments, previously done by hand; NVQLink will let GPUs handle the task in near real-time.

Quantum computer and software provider Quantinuum has announced that it has NVQLink on its roadmap. At GTC DC, the company demonstrated quantum error correction on the Helios QPU.

NVIDIA's Quantum Ambition

NVIDIA’s quantum work goes into software too. CUDA Quantum, aka CUDA-Q, was launched in 2022 as a platform for hybrid quantum-classical computing, analogous to the CUDA libraries for GPU applications. It provides the foundation for this next generation of heterogeneous computing.

NVIDIA’s interest in quantum computing makes sense, given that the company describes itself as a driver of next-generation computing, not just AI. Quantum computing, once it gets to production-level workloads, will probably always work alongside a supercomputer, because quantum computers aren’t good at the massive calculations that classical computing does. Expect NVIDIA to continue emphasizing that kind of heterogenous computing.