This week’s unveiling of the Blackwell GPU may demonstrate Nvidia chief Jensen Huang’s constant quest for bigger and better GPUs, but the company’s existing GPUs are nothing to sleep on. Case in point: The Nvidia H100 GPU family has been in production since 2022 and there is still plenty of demand for it to support cutting-edge applications, such as supercomputers being used for quantum computing research.
While Blackwell may have gotten more attention during Nvidia’s GTC event this week, the company also announced two new quantum-focused supercomputer projects involving the H100. The first is Japan’s new ABCI-Q supercomputer, which is being built now by Fujitsu at at an advanced technology R&D hub in Japan. ABCI-Q will rely on more than 2,000 Nvidia H100 Tensor Core GPUs in more than 500 nodes interconnected by Nvidia’s Quantum-2 InfiniBand technology, according to Tim Costa, director of HPC and quantum computing at Nvidia.
ABCI-Q is designed to advance Japan’s national quantum computing initiative, and with the help of Nvidia’s CUDA-Q open-source hybrid quantum computing platform it will be used in high-fidelity quantum simulations for research across many different industries.
Such hybrid supercomputer use cases are nothing new for Nvidia, as the company last month said that Grace Hopper Superchips and the CUDA-Q platform would be used to power quantum simulations at Australia’s Pawsey Supercomputing Research Centre. However, Costa called ABCI-Q “the largest single deployment dedicated to quantum research to date.”
The second quantum-focused supercomputer project Nvidia announced this week, though smaller than the one in Japan, still involves a significant number of H100 GPUs. In Denmark, the Novo Nordisk Foundation is deploying a DGX SuperPOD, which involves multiple DGX systems, each of which has eight H100 GPUs, as part of its 12-year plan to build up a quantum computing infrastructure.
This sort of GPU-powered hybrid quantum-classical computing is becoming more common, and Costa said such deployments will continue to involve larger numbers of GPUs. “The community has realized that in order to address challenges in designing qubits with better fidelity, figuring out how to build up a larger scale of qubits to work on HPC quantum integration, to design algorithms for error correction, to just discover algorithms with exponential speed up, as well as to develop all the developer tools that we need in order to do useful quantum computing–all of this requires GPU supercomputing as the tool to advance that research.”