Simulation and HPC

The workload

High performance computing on GPUs covers computational fluid dynamics, structural analysis, molecular dynamics, quantum chemistry, seismic processing, weather and climate, and the fast growing class of physics informed AI models that replace parts of a simulation with a learned surrogate. These codes have run on GPUs for a decade, and the mature ones scale across nodes with MPI over InfiniBand. The newer AI surrogates and foundation models for science look more like LLM training and are sized the same way.

HPC differs from AI in two respects that change the hardware. Many codes need FP64, double precision, which the data center GPUs provide and workstation GPUs mostly do not. And many codes are bound by memory bandwidth and by latency between nodes rather than by raw compute, so the fabric and the memory system matter as much as the GPU count.

What it needs from the hardware

FP64 throughput for the classical codes, which means HGX class GPUs or H200 NVL rather than RTX PRO cards. High memory bandwidth, where HBM3e is the point of the exercise. A low latency fabric: InfiniBand remains the preferred choice for MPI heavy codes, and 400 or 800 Gb/s per node with full non blocking topology keeps strong scaling honest. Plenty of host memory, because pre and post processing, meshing and I/O run on the CPUs, and a parallel file system or at least fast shared NVMe for the results, which in HPC are often larger than the inputs.

Where AI surrogates are part of the pipeline, the same node needs the Tensor Core performance and lower precision formats that training uses, which the Blackwell HGX platforms provide alongside FP64.

Recommended systems

Workstation or small server

For code development, small meshes and teaching, a 4 GPU 4U PCIe server with H200 NVL gives FP64 and HBM in a departmental rack. Where FP64 is not needed, a 4 GPU AI workstation with RTX PRO 6000 Blackwell runs single precision codes and AI surrogates on a desk.

PCIe GPU server

The 8 GPU PCIe server with H200 NVL is the departmental HPC node: eight HBM GPUs with FP64, paired over NVLink bridges, in a chassis that runs on air in a standard high density rack. It covers most single node simulation and is a good citizen in a mixed AI and HPC queue.

HGX node or cluster

The 8 GPU HGX H200 AI server and the 8 GPU HGX B200 AI server are the building blocks of a GPU HPC cluster, with NVSwitch inside the node and an 800 Gb/s InfiniBand rail per GPU between nodes. A 16 node AI pod with a non blocking fabric is a sensible center scale system that runs MPI codes and AI training on the same queue; the HGX B300 is the choice when AI surrogates dominate the mix.

Sizing notes

Match the GPU to the code: check whether your solvers use FP64 and whether the vendor's GPU port scales across nodes, because that decides HGX versus PCIe and InfiniBand versus Ethernet. Plan host memory generously, 240 GB per GPU or more where meshing and post processing are heavy. Give the cluster a shared file system with at least 1 TB usable per GPU and bandwidth to write results from every node at once. Keep the fabric non blocking and keep storage traffic off the compute rails. Our HGX versus PCIe guide covers the interconnect choice in more detail.

Next step

Tell us the codes, the precision they use, the typical problem size and the number of users, and we will propose nodes, fabric and storage that scale together and quote it within one business day. We work with university procurement and grant timelines and can quote in the form your institution requires.