The Age of AI Will See Supercomputing Driven by Networking and Accelerated Computing
In order to demonstrate the quick advancement of AI supercomputing worldwide, NVIDIA presented a wide range of innovations at SC25 in the areas of accelerated computing, next-generation networking, and quantum systems.
NVIDIA’s general manager and vice president of accelerated computing, Ian Buck, gave a special speech on the occasion. New AI physics models, quantum jumps through NVQLink, NVIDIA Quantum-X Photonics InfiniBand CPO networking switches, and the NVIDIA BlueField-4 DPU were among the announcements.
You can also read NVIDIA Bridges AI and Quantum with NVentures Newest Bets
NVIDIA and RIKEN Collaboration Advances Japan’s Scientific Frontiers
In order to construct two new GPU-accelerated supercomputers, NVIDIA is collaborating with RIKEN, the top national research institute in Japan. The purpose of these systems is to further establish Japan as a leader in quantum computing and AI for science.
The new infrastructure will strengthen Japan’s secure domestic infrastructure and sovereign AI agenda by connecting 2,140 NVIDIA Blackwell GPUs in total via the GB200 NVL4 platform and NVIDIA Quantum-X800 InfiniBand networking.
Included in the two new systems are:
- AI for Science System: This system uses 1,600 Blackwell GPUs to support research in climate and weather forecasting, materials science, life sciences, manufacturing, and laboratory automation, among other areas.
- Quantum computing system: 540 Blackwell GPUs are used to speed up quantum-classical techniques, hybrid simulation, and quantum algorithms.
It is planned for these supercomputers to go online in the spring of 2026.
RIKEN’s current efforts to codesign FugakuNEXT, the replacement for the Fugaku supercomputer, alongside Fujitsu and NVIDIA are being expanded upon by this partnership. Production-level quantum computers are anticipated to be integrated by 2030, while FugakuNEXT is anticipated to provide 100x better application performance. Codesigning and developing FugakuNEXT’s hardware, software, and applications will be done through the new GPU-accelerated supercomputers.
In order to establish one of the top unified platforms for AI, quantum, and high-performance computing globally, the RIKEN Centre for Computational Science’s director, Satoshi Matsuoka, said that the NVIDIA GB200 NVL4 accelerated computing platform’s integration with their next-generation supercomputers is a significant step forward for Japan’s scientific infrastructure.
You can also read RIKEN And SoftBank Announce 21 JHPC-Quantum Program
Linking Quantum and Classical Computing with NVQLink
NVQLink is a worldwide interconnect that is being widely used to connect quantum processors and accelerated computing at over a dozen of the world’s leading scientific computing centers. The next generation of quantum GPU, CPU GPU supercomputers is being developed by NVIDIA in collaboration with supercomputing centers throughout the world, according to Ian Buck.
NVQLink is an open architecture that powers large-scale processes using the CUDA-Q software platform by connecting quantum processors and NVIDIA GPUs. Supercomputing facilities are able to integrate a wide variety of quantum processors because to this crucial connection. The AI performance it provides is 40 petaflops with FP4 accuracy.
Among the notable technical accomplishments made using NVQLink are:
- Quantinuum, a firm that specializes in quantum computing, achieved the first real-time decoding of scalable qLDPC quantum error-correction codes by integrating its new Helios QPU with NVIDIA GPUs through NVQLink.
- Using NVQLink correction, the system’s fidelity was about 99%, while it was only about 95% without it.
- The system’s reaction time of 60 microseconds was 16 times faster than Helios’ 1-millisecond criterion. According to a different source, a decoder implementation got a reaction time of 67 microseconds, which was 32 times faster than Helios’ two-millisecond criterion.
By offering a universal bridge between quantum and conventional hardware, NVQLink enables researchers and developers to create real-time quantum-GPU processes, hybrid applications, and scalable error correction.
Adopting facilities worldwide include:
- Asia-Pacific includes the G-QuAT (AIST) and RIKEN Centre for Computational Science in Japan, the Pawsey Supercomputing Research Centre in Australia, the NCHC in Taiwan, the KISTI in Korea, and the National Quantum Computing Hub in Singapore.
- Europe/Middle East: Germany’s Jülich Supercomputing Centre (JSC), Saudi Arabia’s KAUST, the Czech Republic’s IT4I, Italy’s CINECA, Denmark’s DCAI, France’s GENCI, and Poland’s PCSS.
- MIT Lincoln, NERSC, Oak Ridge, Pacific Northwest, Sandia, Lawrence Berkeley, Fermi, Brookhaven, and Los Alamos are some of the top national laboratories in the United States.
You can also read IonQ Error correcting codes Will Improve quantum computing
Photonics Networking and BlueField-4 Drive AI Factories
AI factories’ operating system is powered by the BlueField-4 DPU, which NVIDIA demonstrated. To allow CPUs and GPUs to focus on compute-intensive tasks, the BlueField-4 DPUs offload, accelerate, and isolate essential data center operations, such as networking, storage, and security. With NVIDIA ConnectX-9 networking and a 64-core NVIDIA Grace CPU, BlueField-4 delivers high performance, efficiency, and zero-trust security at scale.
In order to reinvent performance and efficiency for AI and scientific workloads, leading storage innovators such as DDN, VAST Data, and WEKA are implementing BlueField-4. As an example, WEKA is implementing its NeuralMesh architecture on BlueField-4, which uses the DPU to operate storage services.
TACC, Lambda, and CoreWeave intend to incorporate NVIDIA Quantum-X Photonics InfiniBand CPO (Co-Packaged Optics) networking switches into next-generation systems as early as the following year. These switches allow supercomputing facilities and AI factories to significantly lower operating costs and energy usage. Through the direct integration of optics onto the switch and the removal of conventional pluggable transceivers, which are frequently the source of task runtime problems, NVIDIA Photonics switch systems:
- Increase power efficiency by 3.5 times.
- Perform with ten times the resilience of before.
- Make apps run five times longer without any interruptions.
According to Peter Salanki, co-founder and chief technology officer at CoreWeave, NVIDIA Quantum X Photonics is helping to improve power economy and reliability, which are essential for handling large AI workloads at scale.
You can also read IBM Quantum System Two Co-Deploys with Fugaku Architecture
Fresh Software Tools and Systems
A number of new hardware and software solutions were also unveiled by NVIDIA to speed up physics and AI simulations:
- NVIDIA started selling the DGX Spark, a machine that is billed as the smallest AI supercomputer in the world. 128GB of unified memory and a petaflop of AI performance are packed into this desktop compact factor. The Grace Blackwell architecture serves as its foundation, enabling developers to fine-tune models locally and perform inference on models with up to 200 billion parameters.
- NVIDIA Apollo: For AI Physics, a family of open models was introduced at SC25. Applied Materials, Cadence, Siemens, and Synopsys are among the industry leaders that are using these models to simulate and speed up design processes in a variety of domains, such as weather, semiconductors, and computational fluid dynamics. In Apollo, domain-specific knowledge is combined with machine learning frameworks such as diffusion algorithms, transformers, and neural operators.
- NVIDIA Warp: An open-source Python framework specifically created to provide up to 245x GPU acceleration for AI and computational physics application. Warp makes high-performance simulation workflow development easier by providing the accessibility of Python with performance on par with native CUDA code. Among the companies using NVIDIA Warp are Siemens, Neural Concept, and Luminary Cloud.
You can also read NVIDIA CUDA-X libraries power quantum with QuTiP-cuQuantum
Arm Adopts NVLink Fusion
The high-bandwidth, coherent interconnect that was first introduced with Grace Blackwell, NVIDIA NVLink Fusion, is being added to Arm’s Neoverse architecture. With the goal of eliminating memory and bandwidth constraints that restrict AI performance, NVLink Fusion unifies CPUs, GPUs, and accelerators into a single, cohesive rack-scale architecture. This integration guarantees smooth data transfer between Arm-based CPUs and accelerators and is connected via Arm’s AMBA CHI C2C protocol.
This partnership establishes a new benchmark for AI infrastructure, allowing ecosystem participants to create unique, energy-efficient solutions. Ian Buck has verified that “Folks building their own ARM CPU, or using an Arm IP can actually have access to NVLink Fusion, be able to connect that ARM CPU to an Nvidia GPU or to the rest of the NVLink ecosystem”.
The Comparative Analysis
Comparatively speaking, NVQLink’s implementation to integrate quantum processors with GPU supercomputers is comparable to constructing a universal translator over a fast fiber-optic connection. The GPU communicates in huge parallelism (classical computing), while the quantum processor communicates in quantum physics, the language of nature. The crucial, ultra-low latency connection and translation layer (CUDA-Q) APIs that NVQLink offers allow these two radically different systems to work together in real-time on challenging tasks like error correction, facilitating quicker and more accurate scientific discoveries.
