Nvidia arrived at the SC22 supercomputing conference in Dallas this week with numerous announcements, including news of partners adopting its Quantum-2 Infiniband high-speed networking solution and H100 Tensor Core GPUs.
Microsoft Azure was revealed as the first customer for the Infiniband offering, which was announced earlier this year.
“The future of transformative enterprise technologies such as AI and HPC is in next-generation cloud platforms like Microsoft Azure, where innovators have the opportunity to deliver a new era of technological breakthroughs,” said Nidhi Chappell, general manager of Azure AI Infrastructure at Microsoft. “The NVIDIA Quantum-2 InfiniBand networking platform equips Azure with the throughput capabilities of a world-class supercomputing center, available at cloud scale and on demand, and allows researchers and scientists using Azure to achieve their life’s work.”
Meanwhile, Dell, Lenovo, and HPE were among at least 10 companies that were announced as partners offering server products employing the H100. Dell’s PowerEdge XE9680, which targets AI and high-performance workloads, uses HGX platform technology, and Rajesh Pohani, vice president of portfolio and product management for PowerEdge, HPC and Core Compute at Dell Technologies, said the combination of PowerEdge servers with Hopper GPUs would allow its customers “to push the boundaries and make possible new discoveries across industries and institutions.”
Nvidia also announced updates to different software libraries that it has developed to accelerate different kinds of computing projects. These include:
-
The company’s CUDA libraries now include a multi-node, multi-GPU Eigensolver enabling unprecedented scale and performance for leading HPC applications like VASP, a package for first-principles quantum mechanical calculations.
-
The cuQuantum software development kit for accelerating quantum computing workflows now supports approximate tensor network methods.
-
DOCA, the open cloud SDK and acceleration framework for Nvidia BlueField DPUs, includes advanced programmability, security and functionality to support new storage use cases.
Nvidia also said at SC22 that its Omniverse platform for creating metaverse applications and digital twins now connects to scientific computing visualization software, such as such as Kitware’s ParaView, an application for visualization; NVIDIA IndeX® for volumetric rendering; NVIDIA Modulus for developing physics-ML models; and NeuralVDB for large-scale sparse volumetric data representation.
Nvidia also announced fully real-time scientific and industrial digital twins for the high performance computing community, enabled by NVIDIA OVX computing system aimed simulation environments. Also, Omniverse now supports batch workloads that AI and HPC researchers, scientists and engineers can run on their existing A100 or H100 systems — including rendering videos and images or generating synthetic 3D data.