NVIDIA Supercharges Science with GH200 and GB200 Superchips
Zach Anderson Mar 18, 2025 14:03
NVIDIA introduces CUDA-X libraries powered by GH200 and GB200 superchips, promising up to 11x faster computational engineering tools and 5x larger calculations.
NVIDIA has unveiled its latest advancements in computational engineering with the introduction of CUDA-X libraries powered by the GH200 and GB200 superchips. These innovations, announced at the NVIDIA GTC global AI conference, promise to revolutionize the field by offering substantial improvements in speed and computational capacity, according to NVIDIA.
Enhancing Computational Engineering
The new CUDA-X libraries enable developers to leverage enhanced integration between CPU and GPU resources, leading to up to 11x faster computational engineering tools and the ability to perform calculations that are 5x larger than those possible with traditional architectures. This advancement is poised to significantly enhance workflows in engineering simulation and design optimization.
Since the release of CUDA in 2006, NVIDIA has developed over 900 domain-specific CUDA-X libraries and AI models, facilitating accelerated computing across various scientific fields. These libraries now extend their impact to disciplines such as astronomy, particle physics, automotive, and aerospace engineering.
Accelerating Engineering Solvers
The NVIDIA cuDSS library, integrated with the new superchip architectures, provides significant performance improvements for solving large engineering simulation problems. By utilizing the Grace GPU memory and NVLink-C2C interconnect, cuDSS can handle large matrices that typically exceed device memory. This capability allows users to tackle extremely large problems efficiently.
Leading engineering software providers, such as Ansys and Altair, have integrated cuDSS into their solvers, achieving remarkable speed improvements. Ansys's HFSS solver, for instance, has seen up to an 11x performance boost in its matrix solver, while Altair's OptiStruct has experienced substantial acceleration in finite element analysis workloads.
Scaling Simulations and Quantum Research
The GH200 and GB200 architectures also enable scaling of memory-limited applications by providing CPU and GPU memory coherency. This feature is particularly beneficial for complex engineering simulations requiring massive data processing. Autodesk, for example, has successfully conducted simulations involving up to 48 billion cells using eight GH200 nodes, a significant leap in capability.
In the realm of quantum computing, the NVIDIA cuQuantum library accelerates simulations of quantum algorithms, which are critical for developing new quantum computing frameworks. The GH200 system offers up to 3x faster performance on quantum computing benchmarks compared to previous systems, effectively supporting large-scale quantum simulations.
For more information on these advancements, visit the NVIDIA blog.
Image source: Shutterstock.jpg)