Memory channels matter more than cores
A CFD calculation moves huge amounts of data between memory and processor. Each memory channel is a lane it arrives through. When there are many cores per channel, they are left waiting: beyond about four to six cores per channel, adding cores speeds things up very little.
That is why a 32-core processor with eight channels can solve as fast as, or faster than, a 64-core one with four. Before looking at cores, look at channels.
The platforms, by channels
| Platform | Channels | Cores it uses well | Machine | Price |
|---|---|---|---|---|
| Ryzen 9 | 2 | 8 to 12 | Pro, 64 GB ECC | €6,770 |
| Threadripper | 4 | 16 to 24 | Advanced, 64 GB registered ECC | €10,105 |
| Threadripper PRO | 8 | 32 to 48 | Enterprise, 256 GB registered ECC | €30,435 |
| EPYC 9005, in a server | 12 | 48 to 72 | EPYC 9005 server platform, from | €6,152 |
Fill every channel with identical modules. A Threadripper PRO with four modules instead of eight works at half its bandwidth.
How much memory: by millions of cells
Memory is set by the mesh. As a guide, allow between 1 and 4 GB per million cells, depending on the software, the precision and the physical models. A 20-million-cell mesh needs in the order of 64 to 128 GB.
If memory runs out, the calculation uses the drive and goes from hours to days. It is the one component where falling short cannot be fixed on the fly.
Per-core licences: do not pay for slow cores
Ansys Fluent and STAR-CCM+ are licensed by parallel cores. With a licence for 16 or 32 cores, what matters is that those cores are fast and have channels to spare. A processor with fewer cores and higher frequency gets more out of the same licence.
OpenFOAM is free: there it does pay to fill every core the channels can feed.
The graphics card: more and more, but not always
Fluent and STAR-CCM+ have solvers that compute on the graphics card, and in the cases they support they are much faster. The limit is graphics memory: the whole mesh has to fit. That is why cards with 32 to 96 GB are used, or several.
For meshing and viewing results, the graphics card matters little: a mid-range one is enough.
Drives: results take up space
A transient run saves a result for each step and fills terabytes in days. Use a fast NVMe drive to work on and high-capacity drives, or a file server, to store. And decide from the start what is kept and what is deleted.
Related hardware and guides
Frequently asked questions
How many cores do I need for CFD?
As many as your memory channels can feed and your licence allows. As a rule, four to six cores per channel: 8 to 12 on a Ryzen 9, 16 to 24 on a Threadripper and 32 to 48 on a Threadripper PRO.
Threadripper or Threadripper PRO?
Threadripper has four memory channels and the PRO has eight. For meshes above about 20 million cells, the PRO clearly performs better even with the same number of cores.
Is computing on the graphics card worth it?
If your software and physical models support it and the mesh fits in graphics memory, yes: the difference is large. If not, the money goes further on memory channels and fast cores.
Workstation or compute server?
To work by day and compute by night, a workstation. If calculations last days or are shared by several people, a rack server with a job queue and remote management, plus a lighter seat to prepare and view.
Shall we size your workstation?
Tell us which software you use and the size of your projects. We will propose the processor, memory, GPU and drives, with a price, within 12 working hours.