Intel P +E or AMD for large (2000) Monte Carlo simulations

I'm about to build a new pc with running large Monte Carlo simulation in mind. I know that matlab doesn't benefit from threading. Am I better off with an Intel 14700k 20core (8P+12E) cpu, or an AMD 7950x 16core cpu (16P) cpu.

I'm concerned that the intel chip E cores can't perform equally well as the P cores, so that different Monte Carlo runs won't finish about the same time.

On the other hand, by running bench(0), I see that AMD chips are really bad at linear algebra, which is another concern to me.

A related q&a to this post

https://www.mathworks.com/matlabcentral/answers/2062522-how-does-parallel-computing-toolbox-handle-performance-and-efficiency-core-usage

Answer 1

[Qiang Li]

I went with the intel 14900k. For me it seemed to be the safe bet at that moment.

https://blogs.mathworks.com/matlab/2022/07/13/linear-algebra-in-matlab-trying-out-amds-aocl/#:~:text=AOCL%20is%20optionally%20available%20in%20MATLAB%20from%20R2022a&text=As%20such%2C%20R2022a%20continues%20to,has%20passed%20MathWorks%20qualification%20testing.

After using the intel chip for over a month mostly doing parallel computing, it becomes clear that the E cores are holding the P cores back. Monitering the system resources, the P cores always finish the job first and will idle waiting for the E cores to catch up.

[Yunxiao]

Hi, my new PC is still under delivery and I am worried about the same issue, too. Could you please tell whether the relative waiting time for P cores would not be that serious if the number of parallel loops is large? The process is like below:

Suppose there are 8 P cores and 16 E cores, and I called 24 workers for parallel computation (each worker is based on a physical core). Also suppose there are last 24 parallel loops left, and the 24 workers are working on it. since 8 P cores are faster and they finish 8 loops first, then they would wait for 16 E cores to finish. Then if the number of parallel loops to be computed is quite large, then the relative waiting time for P cores would not be that serious, right?

[Qiang Li]

Like you said, when the loop is large enough, I've tried 300 so far, all cores will stay active most of the time. But I've also observed when the p cores would idle for half an hour for a simulation that took three hours. It could be caused by my implementation. I guess you also went with the i9 chip, mine will stay 5.5ghz on the p cores and 4.3 on the e cores over prolonged period of time with a 0.1 undervolt.

edit: I ran some 100 parfor loops today. Here're some cpu status. Hard to tell if the active P/E-core count reflects the actual workload on them.