It is great to hear that you get good results and benefit from it!

In general the performance of Fleur depends on the hardware as well as on the software infrastructure of compilers and libraries. The more exotic your hardware is the more problematic is getting a good software infrastructure that works and gives good performance. Of course, we always work on making Fleur more performant on every type of hardware it is used on. This implies that the best hardware may come with problems now but has large benefits in a few years. But I guess you want to have a solution now and not sometime in the future.

1. I think for a normal Fleur user computations on GPU hardware should be avoided at the moment. This is because the available software infrastructure still needs some development to make flawless installation and usage of the code possible. Also we probably have to tune a few more code sections to such setups.

2. The next choice is Intel or AMD hardware. Fleur works on both but there are differences. Traditionally many of the used software libraries and also Fleur itself was mainly used on Intel hardware. This means that it is easy to get good performance out of such a setup. For AMD hardware the choice of software libraries, compilers, and compiler options again is critical to get good performance. But the issue is much smaller than for GPU hardware. Our inhouse cluster bought last year actually uses AMD EPYC CPUs. You should also know that the AMD processors get their performance advantage over Intel processors mainly by having more cores. This means that you have to take care to choose good parallelization schemes on such hardware to get the best out of it.

3. For our inhouse cluster we have 128 GB RAM for 32 CPU cores on each node. That should be a good RAM/cores ratio.