At SC25, NVIDIA unveiled advances across NVIDIA BlueField DPUs, next-generation networking, quantum computing, national research, AI physics and more — as accelerated systems drive the next chapter in AI supercomputing.
NVIDIA also highlighted storage innovations powered by the NVIDIA BlueField-4 data processing unit, part of the full-stack BlueField platform that accelerates gigascale AI infrastructure.
More details also came on NVIDIA Quantum-X Photonics InfiniBand CPO networking switches — enabling AI factories to drastically reduce energy consumption and operational costs — including that TACC, Lambda and CoreWeave plan to integrate them.
Last month, NVIDIA began shipping DGX Spark, the world’s smallest AI supercomputer. DGX Spark packs a petaflop of AI performance and 128GB of unified memory into a desktop form factor, enabling developers to run inference on models up to 200 billion parameters and fine-tune models locally. Built on the Grace Blackwell architecture, it integrates NVIDIA GPUs, CPUs, networking, CUDA libraries and the full NVIDIA AI software stack.
DGX Spark’s unified memory and NVIDIA NVLink-C2C deliver 5x the bandwidth of PCIe Gen5, enabling faster GPU-CPU data exchange. This boosts training efficiency for large models, reduces latency and supports seamless fine-tuning workflows — all within a desktop form factor.
NVIDIA Apollo, a family of open models for AI Physics, was also introduced at SC25. Applied Materials, Cadence, LAM Research, Luminary Cloud, KLA, PhysicsX, Rescale, Siemens and Synopsys are among the industry leaders adopting these open models to simulate and accelerate their design processes in a broad range of fields — electronic device automation and semiconductors, computational fluid dynamics, structural mechanics, electromagnetics, weather and more.
The family of open models harness the latest developments in AI physics, incorporating best-in-class machine learning architectures, such as neural operators, transformers and diffusion methods, with domain-specific knowledge. Apollo will provide pretrained checkpoints and reference workflows for training, inference and benchmarking, allowing developers to integrate and customize the models for their specific needs.
