Hardware Recommendations

Last Update: 2025-03-24

Table of contents

General rule of thumb is to use one core per 50 000 elements, and to run your first simulation with a high memory core type, if you are unsure about the memory usage. Since running out-of-core memory is not supported with CloudConnect, running out of memory will cause the simulation to crash. After the first run core type can be changed to a low memory type, if extra memory is not needed. It is also possible to limit the number of cores the solver uses to get more memory, for example if you are limited by the number of licenses available.

General simulations with low memory demand

Jacinth

  • Intel Xeon (Sapphire Rapids) @ 3.5 GHz
  • 4 GB memory per core
  • 96 cores per node

Simulations with high memory demand

Malachite

  • 3rd generation Intel Xeon Scalable processors (Ice Lake) @ 3.5 GHz
  • 16 GB memory per core
  • 64 cores per node

GPU accelerated simulations

Mallorn

  • NVIDIA A100 80GB
  • AMD EPYC 7V13 w/ NVIDIA Ampere A100 @ 2.5 GHz
  • 9.17 GB memory per core
  • 4 GPUs per node
  • 48 cores per node

Core types shown in CloudConnect can be set using the Core Configurator

Monitoring Hardware Usage and Scaling Efficiency

In Ansys Mechanical you can track hardware usage under Solution → Solution Information → Solution Output → Solution History. There you can also track how well the solver is able to distribute the simulation with the Element Load Balance Ratio. For running jobs, you can track hardware usage report from Rescale Job Analyzer.

Element Load Balance Ratio tells the ratio of CPU time spent most utilized core and the least utilized core, or in other words the slope of cores per simulation. If the value is close to one, the solver is able to distribute the solution very well and you could increase the number of cores used. When the value starts getting close to 1.5-2.0, the speed up from just adding more cores will start to diminish.

When using large amount of cores, it is also a good idea to keep make sure the simulation can be solved on a single hardware node, to avoid the communication delay between nodes. If your simulation scales well on multiple hardware nodes, remember to use a core type with high interconnect speed.

A general recommendation is to use as many cores as reasonably possible to reduce the simulation time. The rule of thumb is hardware costs are very low, but engineering time and solver licenses are expensive, so minimizing the time spent waiting for a solution is key.

GPU Acceleration

Using a GPU for acceleration is enticing due to the fact a single GPU only takes one Ansys HPC license to run and the type of the GPU has no effect either. Because of this it is recommended to use a as powerful GPU as possible for best return on the HPC license.

GPU acceleration is one key focus in solver development, but it does not yet fit all analysis types. Best results for GPU acceleration can be gained in highly nonlinear analyses. For harmonic and modal analyses GPU acceleration is not recommended.

Ansys Help: Supported GPU Hardware, Analysis Types, and Features