AI Trends

NVIDIA Releases Rubin HPC GPU to Accelerate Agentic Applications in Science and Engineering

cnBetaUpdated Nov 17 01:42

AI summary

NVIDIA has released the Rubin HPC GPU for science, engineering, and data analysis, with each GPU equipped with more than 50,000 CUDA cores, setting a record for the number of CUDA cores. The GPU features sparse data acceleration capabilities, adopts seventh-generation NVLink interconnect, and uses a balanced design in which data movement and computation scale in sync to accelerate simulation workloads such as sparse matrix operations, stencil computations, sorting, and reductions. Rubin HPC is expected to be officially launched in 2028, and the CUDA-X software libraries will provide optimized implementations for it.

Body · AI translation

NVIDIA says that as AI agents help scientists and engineers explore more candidate materials, molecules, and design schemes, these schemes must rely on simulation to verify and guide subsequent decisions. Rubin HPC is designed precisely to meet this kind of large-scale simulation demand, with the goal of shortening the path from exploration to scientific discovery and industrial design implementation.

The Rubin HPC GPU is equipped with more than 50,000 CUDA cores, setting a record for the number of CUDA cores in a single GPU. It also features sparse data acceleration capabilities, which can continuously supply data to the cores during computation; at the same time, it adopts seventh-generation NVIDIA NVLink interconnect technology to improve communication efficiency between GPUs.

Ian Buck, NVIDIA's vice president of hyperscale and high-performance computing, said that for agents handling scientific and engineering tasks, faster simulation means faster discovery. Rubin HPC can accelerate the simulations that agents use to test ideas and determine next steps, helping researchers and engineers explore more possibilities and bringing breakthroughs sooner.

From climate models and engineering simulations to genomic analysis and financial risk calculations, scientific computing workloads often share similar characteristics: applications need to extract scattered numerical values from massive datasets, complete computations, and then execute them repeatedly at extremely large scales. To this end, Rubin HPC adopts a balanced design in which data movement and computing capability scale in sync, rather than simply stacking computing power.

This design ensures that the GPU's massive CUDA cores continuously receive data, reducing wait times and devoting more resources to actual computation. NVIDIA claims that Rubin HPC can accelerate recurring computational patterns in simulation workloads such as sparse matrix operations, stencil computations, sorting, and reductions, and these patterns are widely present in data analysis, scientific research, and engineering applications.

NVIDIA also emphasized that the CUDA-X software libraries will provide implementations optimized for complex algorithms, covering key tasks in the simulation and analysis fields. When used with Rubin HPC, these libraries can help agents obtain inference results faster, shorten the loop of "proposing hypotheses, running simulations, and optimizing schemes," allow researchers to determine the next direction earlier, and concentrate laboratory resources on the most promising candidate schemes.

Rubin HPC is expected to be officially launched in 2028. NVIDIA positions it as a dedicated high-performance computing GPU for science, engineering, and data analysis, hoping to use higher simulation speeds and stronger data throughput capabilities to drive cross-industry scientific research and industrial innovation.

Source

cnBeta