I have quite seed money to purchase a computer. What is the best choice for faster calculation?

Hello.

I need some help what I should purchase a computer system.

I am not familiar with hardware, especially, high end computer, and I do not have enough time to search and/or developing a computing system.

I usually calculate a electro-magnetic circuit including thousands of R-L-C. and it tooks about a couple months with my PC (i7-9700K, 64 GB RAM, GTX1060, not over-clock)

I have about $ 50,000.

1. Cloud computer
2. Cluster computer
3. Many high-end PCs

How much take the time to calculate the circuit?

[Walter Roberson]

How feasible is it to parallelize your calculations ? How easy is it to vectorize your calculation?

Is it effectively a FEM type of calculation?

[Jaemin Kim]

It's alomst vector calculation. Size of the vector is thousands by thousands (# of RLC).

[Walter Roberson]

Can the problem be sub-divided into problems that can be solved independently, with (relatively) small effort to combine the sections ?

I am concerned about "diminshing returns" -- that after a point, adding more CPUs might not improve performance much due to the time to automatically set up the parallel problems.

Can you extract a representative sample of your calculations and run them with all of your cores, and then use maxNumCompThreads() to tell it to use only half as many cores, and then time again? Better yet would be to automatically test from roughly half of your physical cores up to your full number of cores, timing as you go, and plot the timing result on a log plot, to give you an idea of how linear the calculation is in the number of cores.

There are some problems where having many many but slower cores is fine, but there are other problems where you only need a moderate number of cores but you want those cores to be as fast as practical.

Also: I see you have a GTX1060. Is that being actively used in your calculation at the moment? If so are you using double precision calculations or single precision? If you are currently using double precision, would it be reasonable to move to single precision?

[Jaemin Kim]

Thank you very much.

Your comment would be very helpful to reduce the time.

I will try calculation of a sample with maxNumCompThreads() function.

Actually, my matrix is like

m and k is much smaller than n. (about m & k ~ 30, n ~ 3000)

We need to calculate very precise digit because of huge effection from initial error.

Thank a lot. I will find out the parallel calculation effect.

[Walter Roberson]

If the calculations involve matrix multiplications, then with the algorithms used by BLAS, the graphs I have seen over the years suggest that for the size of matrix you are using, peak effectiveness is probably around 4 cores, with performance improvements still available to 6 cores, but by the time you reach 8 that you are not gaining much.

However, if you are doing calculations such as eigenvalue calculations, then I do not recall having seen relevant performance graphs.

[Jaemin Kim]

I really appreciate your kindness.

My calculation is mainly matrix multiplications.

I don't have knowledge of the algorithms but I think it is the clue where I begin.

Thank you!!!