NVIDIA 发布 Rubin HPC GPU,面向智能体驱动的科学与工程
NVIDIA Unveils Rubin HPC GPU for Agent-Driven Science and Engineering
NVIDIA 在 SC26 上发布 Rubin HPC GPU,面向科学、工程与分析场景,单芯片 CUDA 核心数超过 50,000。该芯片针对智能体生成大量候选材料、分子和设计后需要反复仿真的工作流,在稀疏矩阵、stencil 计算、排序和归约等环节做稀疏数据加速,并搭配 CUDA-X 库和第七代 NVLink。Rubin HPC 属于 Vera Rubin 平台,预计 2028 年可用。
正文
NVIDIA announced Rubin HPC at SC26 on November 16, 2026, describing it as a simulation and data-processing GPU for science, engineering and analytics with a record count of more than 50,000 CUDA cores on one chip.
The chip is aimed at workflows in which agents help scientists and engineers generate more candidate materials, molecules and designs, then rely on simulations to evaluate those ideas and guide subsequent steps. Expanding the pool of candidates increases the simulation workload, making computational feedback part of the pace of discovery and industrial design.
Balancing Compute and Data Movement
NVIDIA identifies a shared computational pattern across climate modeling, engineering simulation, genomic analysis and financial risk. These applications repeatedly calculate results from small sets of values scattered through very large datasets, performing the same basic sequence at enormous scale. Arithmetic capacity alone cannot address that constraint; data must reach the cores quickly enough to keep them occupied.
Rubin HPC’s design gives data movement equal weight to arithmetic capacity. Its sparse-data acceleration is intended to prevent cores from idling while they await information. The targeted routines include sparse-matrix and stencil calculations, sorting and reductions, operations that recur across simulation and analytics. Together, those choices address repeated data-access patterns that can otherwise limit the benefit of a large core count.
“For agents tackling science and engineering, faster simulation means faster discovery,” said Ian Buck, NVIDIA’s vice president of hyperscale and high-performance computing. The design goal is a shorter cycle from hypothesis to simulation to refinement, helping teams decide what to test next and prioritize laboratory work around stronger candidates.
Software and Multi-GPU Context
CUDA-X supplies optimized algorithm implementations that agents can use with Rubin HPC, including routines for simulation and analytics. Its libraries are designed as drop-in components, shifting low-level GPU optimization away from application developers. The documented range covers math, scientific computing, physics, parallel algorithms and data processing, providing prepared building blocks that reduce the need for each team to optimize every workload from scratch.
Data movement also extends beyond one processor. NVLink is NVIDIA’s direct GPU-to-GPU interconnect; NVLink Switch joins multiple links so GPUs can communicate all-to-all across a rack, and it can extend those connections across nodes. Rubin HPC uses seventh-generation NVLink, so its data-movement design also covers communication between GPUs.
Rubin HPC is part of the Vera Rubin platform, which NVIDIA describes as five purpose-built rack-scale systems combined into a multi-rack AI supercomputer for multi-step reasoning and long-context agent workflows. That broader platform is ramping into full production and Vera Rubin-based systems are shipping; Rubin HPC itself is expected to be available in 2028.
For agent-driven research, the practical measure of Rubin HPC will be whether balancing compute, data access, software and multi-GPU communication can shorten the path between an agent proposing a candidate and researchers receiving a simulation result that informs the next experiment or design.
AI 热点