In Rudy's example (Yasara) it's most likely some form of MM or MD, which scales much better than QM and works with GPU, so it's not a good comparison.

As for QM: I'm not that familiar with Orca and it may depend on the specific code, but there are some general points that are similar for all QM programs (based on my experience with two codes similar to orca).

Memory: you need to have some minimum memory to be able to do the calculations (scaling is at around 3N2 with the number of basis functions). For standard calculations (DFT with medium/large basis set) usually 2Gb per core is completely enough. Having more memory doesn't really speed things up.

Clock speed: you want to have as fastest proc as possible, the dependence is probably less than linear (probably more like logN), but you should see a quite large difference between e.g. 2GHz and 3GHz.

No of cores: no QM code is perfectly parallelized (there are some implementations of linear SCF, but not in the most used programs). You can usually get something like 1.7-1.9 speedup when going 1->2 cores, 2.5-3.5 speedup when going 1->4 cores and so on for single point energies (and geometry optimizations). For frequency calcualtions the scaling is usually quite worse. At some point it doesn't make sense to add more cores - it's actually better (more time-efficient) to e.g. run twelve 2-core jobs than two 12-core jobs.

In your case a 3-core 4Gb freq job should probably be faster than 1-core, 12Gb job. The most efficient thing, though, would be to run three 1-core 4Gb jobs at the same time - when compared to running 3-core 4Gb freq jobs one after the other finishes.

I am not sure how the code is written for your calculations in order to determine the best way to speed up the process (e.g., clock speed, number of cores, amount of RAM).

In my case, I am doing molecular dynamics simulations in YASARA, and because of the way the program handles the calculations (including balancing between the CPU and the GPU), I have found that the most important parameter is the clock speed of the CPU and secondarily the GPU that I am using. I am using an i7 5960 running at 4.4 GHz, and this gives me faster performance than other machines with more cores and more RAM running slower processors.

Hi Rudy,

Thank you for your response.

Currently the code is written to use a single core (serial processing) and 4 gb of RAM on my i7. The clock speed on the 4790 is 4 GHz. ORCA only allows CPU handling at the moment so a GPU is not involved. From your answer it sounds like more cores and more RAM at only 2 GHz clock speed is not a good option (the AMD Opteron).

So I guess my question boils down to is a single core high RAM (12 gb) setup better than a three core 4gb of RAM per core setup? I.e. How much weighting does RAM amount actually have in these sort of vibrational calculations?

In Rudy's example (Yasara) it's most likely some form of MM or MD, which scales much better than QM and works with GPU, so it's not a good comparison.

As for QM: I'm not that familiar with Orca and it may depend on the specific code, but there are some general points that are similar for all QM programs (based on my experience with two codes similar to orca).

Memory: you need to have some minimum memory to be able to do the calculations (scaling is at around 3N2 with the number of basis functions). For standard calculations (DFT with medium/large basis set) usually 2Gb per core is completely enough. Having more memory doesn't really speed things up.

Clock speed: you want to have as fastest proc as possible, the dependence is probably less than linear (probably more like logN), but you should see a quite large difference between e.g. 2GHz and 3GHz.

No of cores: no QM code is perfectly parallelized (there are some implementations of linear SCF, but not in the most used programs). You can usually get something like 1.7-1.9 speedup when going 1->2 cores, 2.5-3.5 speedup when going 1->4 cores and so on for single point energies (and geometry optimizations). For frequency calcualtions the scaling is usually quite worse. At some point it doesn't make sense to add more cores - it's actually better (more time-efficient) to e.g. run twelve 2-core jobs than two 12-core jobs.

In your case a 3-core 4Gb freq job should probably be faster than 1-core, 12Gb job. The most efficient thing, though, would be to run three 1-core 4Gb jobs at the same time - when compared to running 3-core 4Gb freq jobs one after the other finishes.

Dear  Richard,

In my PC I have the same cpu with 8GB. First of all I suggest you to use the 4 cpu available, which will increase at maximum the capacity. Take into account that there are other 4 logic cpu that will be not used, but you can use for other purposes. About memory, RAM it is not a critical point for DFT, which requires a low quantity in general. The method that it is a determinant point is in MP2, which you can improve the velocity with increasing the RAM, more than with increasing number of cpu.

So, for DFT I suggest you to use 4 cpu and 4-6 GB RAM. Other thing that is very useful is the tool of the motherboard which is a critical component too. Which one do you have and do you have the controlling software for tunn? For example, I can tunn the power supply, frequency of cpu, etc.

Dear Bartosz and Joaquim,

That is exactly what I wanted to know, thank you both!

Joaquim, my motherboard is a Asus Maximus Impact VII with full controlling software, so I have no problems on the motherboard front.

Richard

Dear Richard,

Then, in my observations, the cpu current is more important than cpu load-line. Full for first and half for the other it is enough with a good yield of temperature. Also increase the maximum the velocity of refrigeration is a good idea. Performance around 50-60ºC is quite good and the cpu remains stable enough, but also you can achieve 80ºC but no more than 90ºC.

As for ORCA numerical frequencies, you need to calculate single point for each displacement of each nucleus (resulting in many single points per one freq calculation), so each core is doing its single point alone (there are temporary files where you can see progress of each SCF). For example if you have 4 cores, than you are doing 4 SCF calculations at time (also, those are not entirely occupied cores - just processes, so 4 processes on 4-core machine is perfectly ok). Memory needed for each SCF (unless you system is very big) is quite moderate, hence for average sized system lack of memory is very unprobable (you can estimate how much memory do you need looking at orca_scf process in terminal).

So, in this particular case, the more cores, the faster it would be, I even think, that linearly faster (unless you run out of memory - just check you demandings)

Please, note also, that %maxcore relates to memory devoted per core 

I strongly recommend to use 4 cores and 16 GB of memory. My second recommendation is that you make a simple benchmark, by calculating frequencies in one smaller job with less atoms, and make the same frequency calculation with: a) 1 core (4 GB mem), b) 2 cores (8 GB mem) and c) 4 cores (16 GB mem): hence all the three calculations using the same memory (4 GB/core). Then you calculate the speedups with: S(x)= (walltime in 1 core)/(walltime in x cores). Obviously you will get that S(x) < x, but then you have to estimate which is the optimum number of processors to use according to the ideas above by Bartosz Trzaskowski. I fully agree with him.

Hi German and Oleg,

Thank you for your answers. I have spent some time on this and settled on an optimal configuration as well as also receiving some support on the Orca forums.

As you note Oleg as I am using Orca to simulate Raman spectra I have to use numerical frequency calculations (vibrational frequency simulation of Raman spectra is not supported). Unlike in vibrational frequency calculations, ram usage is low in numerical frequency calculations (I am currently using about 4 gb out of 16 gb for 4 cores, i.e. 1 gb per core). It is my understanding that in numerical frequency calculations %maxcore is not necessary as Orca dynamically assigns memory and I have 16 gb anyway (which I am only using a small % of)

To close this topic and share my findings, as suggested by the other responders above I have found that the number of cores is more crucial and I am currently using all 4 cores of my i7 which has provided a significant speed increase over 3,2 and 1 core calculations. In short calculations (about 30 mins, small molecule), 4 cores is about 1/3 faster than 1 core. In longer calculations the improvement is significantly more, e.g. a ~120 atom molecule takes ~4 days on 4 cores rather than ~2 weeks on a single core.