What changed
Nvidia announced a suite of new tools and hardware at SC26, emphasizing a shift from traditional performance benchmarks to iterative scientific discovery. The company introduced open science models, a CPU-only supercomputing rack, a new HPC-specific GPU, CUDA-X libraries, quantum integration, and co-packaged optics for data centers. These announcements collectively signal a new approach to scientific computing, where AI agents can rapidly iterate through hypothesis, simulation, and refinement.
What this means in practice
Nvidia’s Nemotron for Science package includes the Nemotron 3.5 Super VL model, LitmusBench v0.2, and post-training recipes for building domain-specific science agents. The Vera CPU rack-scale platform, set to ship in early 2027, features 192 Vera CPUs, 300TB of memory, and Quantum-X800 InfiniBand, designed for unaccelerated workloads like climate and molecular modeling. The Rubin HPC GPU, with over 50,000 CUDA cores, is set for a 2028 release and targets sparse-data acceleration. New CUDA-X libraries like cuPhoton, cuPotential, and cuSpa are aimed at applications in astrophysics, drug discovery, and materials science.
Limitations and availability
The Vera CPU rack is available starting in early 2027 through partners like Dell, HPE, Lenovo, and Supermicro. The Rubin HPC GPU is scheduled for release in 2028. Quantum integration with IBM and Fermilab includes a new error-correction decoder and quantum scheduling support in Slurm. Lambda’s co-packaged optics switches connect GB300 NVL72 racks with over 10,000 GPUs. Nvidia’s focus on iterative workflows highlights a shift in how scientific computing is evaluated, with an emphasis on time-to-insight per iteration rather than single-job performance.
