What limits a simulation
A finite element (FEA) or computational fluid dynamics (CFD) solver performs millions of operations on huge matrices. Three things decide how long it takes:
- Cores: the calculation is spread across them, as far as the licence allows.
- Memory bandwidth: data has to reach the cores; if it does not arrive in time, the cores wait.
- Amount of memory: if the model does not fit, the solver writes to disk and everything runs far more slowly.
Memory channels: the hidden bottleneck
Each memory channel provides a fixed throughput. A Ryzen has two, a Threadripper four, a Threadripper PRO eight and EPYC server chips six to twelve. In CFD and many iterative solvers, going from two to eight channels speeds things up more than doubling the cores.
That is why, for serious simulation, you should fill every channel with identical modules: four modules for four channels, eight for eight.
How much memory you need
It depends on the size of the model and the type of solver. Finite element direct solvers are the most demanding; iterative solvers and CFD need less per element, but their models tend to be larger. As a guide:
| Work | Indicative memory |
|---|---|
| Small parts and assemblies, linear analysis | 64 GB |
| Large assemblies, contacts and non-linearity | 128 to 256 GB |
| CFD with several tens of millions of cells | 256 GB or more |
| Multiphysics and very large models | 512 GB or more |
Check the peak memory of your current simulations in the solver log and leave a margin of 30 to 50%: models always grow.
Licences decide how many cores to use
Many simulation programs charge per core, and that changes the sums:
- Ansys includes a few cores with the base licence and charges for the rest through HPC packs.
- Abaqus works with tokens: the more cores, the more tokens.
- COMSOL does not charge per core within a single machine.
If every extra core costs licence fees, a processor with fewer but faster cores and every memory channel filled pays off. If the licence does not limit you, more cores and more channels.
The graphics card: for viewing the model and sometimes for computing
To work with the model, a mid-range professional graphics card with certified drivers is enough, such as an RTX PRO 2000 or an RTX PRO 4000. The question is whether your solver computes on the GPU: some already do (Fluent has a native GPU solver, and Mechanical and Abaqus accelerate part of the calculation), but not all the work benefits.
Before investing in a large card for computing, check which part of your case uses the GPU and whether it needs double precision: today’s graphics cards are very strong in single precision and much weaker in double.
Scratch drive and results
Solvers write huge temporary files, and results take up a lot of space. A dedicated NVMe SSD as a scratch drive, separate from the system, stops the calculation from waiting for the disk; finished results can then move to a server or a NAS with backups.
Three starting points
| Use | Processor | Memory |
|---|---|---|
| Occasional simulation alongside CAD | 16-core Ryzen 9 | 128 GB, 2 channels |
| Daily computation with medium models | Threadripper with 24 to 32 cores | 256 GB, all 4 channels filled |
| CFD and large models | Threadripper PRO | 512 GB or more, all 8 channels filled |
All with a separate NVMe scratch drive and tested for 24 hours at full load, which is exactly what the machine will do when you launch an overnight simulation.
Related hardware and guides
Frequently asked questions
Are more cores always faster in simulation?
No. Beyond a certain point memory bandwidth takes over, and if the licence charges per core, every extra core costs money. The efficient approach is to balance cores, memory channels and licences.
How much memory do I need for Ansys or Abaqus?
It depends on the model and the solver. As a guide, 64 GB for small models, 128 to 256 GB for large assemblies with contacts and 512 GB or more for very large models. Check the peak memory in the solver log and leave a margin.
Is a GeForce card good for simulation?
For viewing the model, yes. For computing, it depends on the solver: some accelerate on the GPU in single precision and others need double precision, where today’s graphics cards are weak. Check it with your own case before buying.
Workstation or compute server?
If one person runs the calculations and wants to see the results at their desk, a workstation. If several people run them or simulations last for days, a rack compute server with more memory channels, shared by the team.
Want to size your compute machine?
Send us the program, the type of analysis and the size of your models: we will propose the processor, memory and drives, with a price, within 12 working hours.