Nvidia Brings Together Quantum And AI For HPC Centers

Nvidia has spent much of the year launching new products and partnerships that aim to ensure it keeps its place atop the wild west that is still the AI market while establishing its place in the emerging quantum computing space, where co-founder and chief executive officer Jensen Huang sees Nvidia being a key infrastructure provider and accelerator.

Most recently, Nvidia’s GTC Washington DC conference in late October featured the company’s typical fire hose of announcements, with Huang and other Nvidia executives unveiling new products like NVQLink, an open high-speed interconnect that links quantum processors to GPUs in supercomputers to create what the company calls “accelerated quantum supercomputers,” BlueField-4, a digital processing unit (DPU) that combines a 64-core variant of the “Grace” CPU and a ConnectX-9 to create an 800 Gb/sec platform for giga-scale AI factories, and open AI models and data.

At the SC25 supercomputing conference this week in St. Louis, much of the focus for Nvidia will be to showcase the partnerships and customer wins coalescing around such technologies, showing the reach it has in the industry and the seemingly ubiquitous nature of its technology. For example, the company announced that more than 80 new Nvidia-powered scientific systems around the world have been introduced over the past year, combining for a total of 4,500 exaflops of AI (meaning the lowest precision on the machines) performance.

In a briefing with reporters, Dion Harris, senior director of HPC and AI infrastructure at Nvidia, talked about the company in 2016 rolling out DGX-1, a highly integrated hardware and software system for AI and accelerated workloads, a first for a vendor known for its GPUs.

“We built DGX1 because we knew that only innovating at the chip level was not sufficient to meet the needs of the coming wave of AI,” Harris said. “Innovating at the system architecture level is not enough. We need to innovate in flops, in compute, and memory, and in systems architecture, but also in scale-up networking, scale-out networking and, of course, software. Nvidia’s platform accelerated computing spans CPUs, GPUs, memory, scale-up networking, scale-out networking, rack-scale architectures, as well as software. We optimized the entire stack from chips to systems to networking to software to applications and we continue to optimize the software, realizing multiple X factors of performance increases throughout the life of the product.”

At the GTC event last month, Huang introduced NVQLink, adding that linking quantum systems to classical supercomputers will address the thorny problem of error correction in quantum systems by leveraging the GPUs in the supercomputers. Adding CUDA-Q, Nvidia’s quantum-classical computing platform, to the mix will let users move beyond error correction and orchestrate quantum devices and AI supercomputers to run quantum GPU calculations.

NVQLink provides 40 petaflops of AI performance at FP4 precision with a GPU-QPU throughput of 400 Gb/sec and a latency of less than four microseconds.

At the time, the vendor noted that nine national laboratories and research centers in the United States – including Los Alamos National Laboratory, Oak Ridge National Laboratory, MIT, Lawrence Berkeley National Laboratory, and the Fermi National Accelerator Laboratory – were adopting NVQLink. At SC25, Nvidia unveiled that more than a dozen supercomputing centers and national research institutions from Asia and Europe are now also embracing the technology to integrate quantum and classical systems.

In Asia, that includes Japan’s Global Research and Development Center for Business by Quantum AI technology, the Korea Institute of Science and Technology Information, and Taiwan’s National Center for High-Performance Computing. European and Middle East centers include Germany’s Jülich Supercomputing Centre, the Poznań Supercomputing and Networking Center in Poland, and the Technology Innovation Institute in the UAE.

A Quantum-Classical Future

“In the future, every supercomputer will draw on quantum processors to expand the problems it can compute, and every quantum processor will rely on a supercomputer to run correctly,” Harris said. “The reason is because computers don’t work alone. They rely on powerful classical processors to manage, stabilize, and interpret quantum operations. GPUs acts as the brain that orchestrates the quantum hardware and perform the heavy computations that quantum processors cannot.”

The integration that’s needed “between quantum processors and GPU supercomputers is extremely demanding,” he added. “Performing tasks like quantum error correction requires a quantum GPU interconnect with microsecond latency and throughput in the hundreds of gigabits per second.”

Harris also noted that, beyond supercomputing and research centers, other quantum companies are adopting NVQLink. He pointed to Quantinuum, which is incorporating the interconnect into its roadmap for future quantum systems and used it in its demonstration of real-time error correction with its new Helios chip.

“This demonstration used extremely low latency algorithms for scalable quantum error correction codes made possible by NVQLink and CUDA-Q,” he said. “These advancements provide Quantinuum with scalable access to GPU supercomputing for their quantum processor, with a total round trip of 67 microseconds, well within the limits that are needed to scale their system.”

Nvidia also announced that Japanese research center Riken is adopting its technology for two new supercomputers, one that will run AI science workloads and the other built for quantum computing. The first will include 1,600 Nvidia Blackwell GPUs and use the GB200 NVL4 platform, a liquid-cooled system for HPC and AI applications that includes two Grace CPUs and four Blackwell GPUs in a single node. The supercomputer will run research jobs in such areas as life sciences, materials science, climate and weather forecasting, and manufacturing.

The system for quantum computing will include 540 Blackwells and also use the GB200 NVL4 platform, all interconnected with Nvidia’s Quantum-x800 InfiniBand network. It will be used for research in quantum algorithms, hybrid simulation, and quantum-classical computing methods.

The SC25 announcement comes two months after Riken, at the FugakuNEXT International Initiative Launch Ceremony in Tokyo, said it planned to work with Fujitsu and Nvidia to co-design FugakuNEXT, the next-generation supercomputer after Fugaku.

In another partnership, Arm announced that CPUs based on its Neoverse design will integrate with AI chips using Nvidia’s NVLink Fusion, an extension of its NVLink high-speed fabric to allow partners to build custom and semi-custom AI infrastructure by integrating their own CPUs, ASICs, and other components with Nvidia GPUs.

This will come as good news to hyperscalers like Google, Amazon, and Microsoft, which are building their own Arm-based chips for their cloud services and now will be able to integrate them with Nvidia GPUs.

In addition, Nvidia announced that the Texas Advanced Computing Center (TACC), Amazon Web Services’ Lambda serverless computing service, and AI cloud computing company CoreWeave will integrate its Quantum-X Photonics InfiniBand COP networking switches, which were introduced in April to save on energy consumption and operational costs.

Meanwhile, HPC storage vendors DDN, VAST Data, and WEKA are adopting BlueField-4 DPUs for a range of jobs, from next-generation AI factories to improving AI pipelines with intelligent data movement to – for WEKA – launching its NeuralMesh AI software-defined storage offering.