How much computer CPU usage will be required for say a 60 atom model to run. So we want CPU requirements versus the number of atoms for VASP

Dear Material project users

In Egypt we are about to invest some money from a project fund to Buy or Rent a high performance computer to calculate the magnetic, electronic, elastic, and magnetocaloric properties of selected rare-earth transition metal compounds to include R2Fe14B, Nd3Co11B4.

The problem is we have to state in our request for this fund how much computer CPU usage will be required for say a 60-atom model to run. Therefore, we want CPU requirements versus the number of atoms for VASP. Also need to estimate how many times these models will need to be executed to perform the field, temperature ranges, first and second order phase transitions, magnetization, specific heat, entropy, and MCE properties i.e. the isothermal change in entropy and the adiabatic change in temperature in a wide range of magnetic fields and in a temperature range up to the Curie temperature of the studied systems.

I am not sure that we can estimate the number of flops; so far, I have not seen this type of information on the internet for VASP. There are tools, but not clear, if there has been a paper on the subject. The issue I see is that we need to estimate how much computer time to ask for, so if we can get an estimate that would be a good idea. If they give us, 200 or 2000 hours will that, be enough to do what we want? How much computer time/gflops do we ask for?

Thank you for your patience.

Dr / Mohammed Said M. Abu-Elmagd

Ph.D. in theoretical physics

Department of Physics, Higher Institute of Engineering, Shourok Academy, Cairo, Egypt.

I’m not sure this is the best forum to ask such a question, but as a rough guess, a 60-atom system might take 2 hours to optimize the atomic geometry on a computing node with 16 CPU cores, so about 30 cpu-hours. Computing the properties you list will require substantially more for each material. However, these estimates can change at least a full order of magnitude in either direction depending on pseudopotentials, accuracy settings, and the performance of the hardware, so it is challenging to guess without running any test calculations first. Please note that the increase of CPU time is not linear with the number of atoms you calculate at once; rather, doubling the number of atoms will at least quadruple running time, perhaps more.

I don’t think you will be able to do much with only 2,000 cpu-hours. When my thesis adviser was doing his own thesis, his advisor told him that he should be able to publish a full theory paper for every 50,000 cpu-hours he spent running VASP. That was about a decade ago, and availability of computational time has only gone up in the mean time - every theory paper I’ve published cost at least 100,000 cpu hours in VASP. You might be able to manage with closer to 20,000 if theory is only a small component of the paper. That is the best guess I can give without seeing any benchmark test calculations for what you are doing.