Scientific computing on GPUs spans molecular dynamics (GROMACS, AMBER), computational fluid dynamics (ANSYS Fluent, OpenFOAM), weather and climate modeling, finite element analysis, and quantum chemistry. These workloads typically demand FP64 throughput, large VRAM, and high memory bandwidth.
FP64 performance separates the field: datacenter GPUs like the A100, H100, and H200 offer FP64 rates close to their FP32 rates (full-rate or half-rate depending on architecture), while consumer GPUs typically run FP64 at 1/32 or 1/64 of FP32 speed. For workloads that don't need FP64 (most ML), this is irrelevant; for CFD and molecular dynamics solving stiff equations, it's the dominant factor.
ECC memory is highly preferred for long-running scientific simulations where a single bit flip can corrupt days of compute. NVLink interconnect enables efficient multi-GPU simulations. CUDA-aware MPI is standard for distributed runs across nodes.