CPU for scientific computing: prioritize core count or speed?

[vardonir]

The programs I mostly use are Matlab, Lumerical, and Comsol, but I'd listen to input from people who use other simulation programs that don't support GPU acceleration. Suppose money (and RAM) is unlimited. Should someone working in scientific simulations get a CPU with more cores or faster single core performance?

 

fwiw, we have two servers in my lab: One running a TR1920x and one with a Xeon E5-1620. But it's not really fair to compare them, considering they're made 5 years apart...

 

 

[TehDwonz]

28 minutes ago, vardonir said:

The programs I mostly use are Matlab, Lumerical, and Comsol, but I'd listen to input from people who use other simulation programs that don't support GPU acceleration. Suppose money (and RAM) is unlimited. Should someone working in scientific simulations get a CPU with more cores or faster single core performance?

 

fwiw, we have two servers in my lab: One running a TR1920x and one with a Xeon E5-1620. But it's not really fair to compare them, considering they're made 5 years apart...

 

 

You'll most likely benefit from a CPU with support for AVX2 - see here: https://en.wikipedia.org/wiki/Advanced_Vector_Extensions#CPUs_with_AVX2

Since AVX/2 generates more heat, the CPU will downclock slightly anyway, so I'd say go for core-count over highest turbo clock.

 

edit: oh, and if your work is important, I'd also lean toward Xeon or Epyc, rather than consumer Intel Core, or AMD Ryzen. ECC RAM being a must.
If it's just hobby level work, then that 64 core X3990 Threadripper would be pretty sweet.

[Aereldor]

If money is unlimited you can have both.

Threadripper 3990x - 64 cores, boosts to 4.3- but not on all cores, I'd wager all-core is closer to 3.5 ghz, even 3.7 stable seems to need upwards of 1.4v.

The 3970x hits an all-core 4.3 GHz on 1.35 volts though, and it's half the price. I don't know how well your programs scale beyond 32 cores, so the 3970x may be a better buy.

[Pixel5]

it really depends what exactly you are calculating, there are things that can be parallelized like data sets where the same calculation has to be done for many different number sets which would benefit greatly from more cores and there can be calculations where you cant continue before the result of the previous calculation is done which would only run on a single core at all times.

 

you rally have to know in detail what you are doing and if it can be parallelized or not and then decide which CPU to go for.