Nvidia announced at ISC High Performance in Hamburg that it is positioning agentic artificial intelligence as the next phase of scientific computing and is shipping a new stack to support it.
The company said the stack links autonomous "agent" workflows, simulators and models so AI co‑scientists can plan experiments, write and run code, and converge simulation, AI and analytics into single workflows.
"AI is shifting from a tool that simply answers questions to an autonomous system that executes complex tasks," Nvidia's Dion Harris told the media.
Nvidia is selling the hardware foundation as the Vera Rubin NVL rack, due to be available in Q4, which packs up to 144 GPUs per rack and delivers 5 petaFLOPS of FP64 performance.
Vera Rubin also raises memory bandwidth 2.8 times versus Blackwell, offers 41 TB of HBM4 per rack and claims three petabytes per second of bandwidth.
Planned deployments include Los Alamos National Laboratory's Mission (2,160 Rubin GPUs and 1,080 Vera CPUs) and Vision (1,298 Rubins and 648 Veras), while Veritas will deploy 576 Rubin GPUs and 288 Vera CPUs.
Nvidia's software layer names ALCHEMI for chemistry and materials discovery, DAQIRI for real‑time instrument inference, and cuPhoton for telescope and camera data processing.
At CERN's ATLAS, Nvidia says DAQIRI's GPU‑accelerated AI trigger uses FPGAs for low‑latency routing and GPUs for deep learning to retain far more collision data.
In tests with 32 Grace Blackwell superchips simulating Rubin Observatory data, Nvidia claimed cuPhoton read images 15,000 times faster and sped signal processing and analysis by up to 8,000 times.
Nvidia also said Europe added 35 new supercomputers using its technology in the past year, including Jupiter, MareNostrum 5, Blue Swan, HammerHAI and CINECA.
