Intel® oneAPI Math Kernel Library
Ask questions and share information with other developers who use Intel® Math Kernel Library.

Seeking advice on new CPU purchase for MKL band matrix LU

hareson
Beginner
289 Views

Hi All

I would like to improve the performance of the LU decomposition speed on my desktop by buying more CPU power. However, I don't know how I should weigh in things like clock speed vs. numbers of cores etc.

The problem: LU decomposition and solving for 1 solution of a NxN band matrix, where N is between 10^5 to 10^6, and the band is about 1000. I use the cgbtrf function for the LU.

CPUs:

I have been thinking of buying two Intel Xeon E5 using the LGA2011 socket for a dual CPU system. However, the price for clock-speed makes the "bang per bucks = total clocks * cores / $US" largest if I choose, say, the 2620 CPU compared to a 2643 or a 2640.

However, I am also aware that the problem will scale linearly by that metric, so I was wondering if anyone could offer some insight into this problem.

Thank you all for your time

Henrik Andresen

0 Kudos
2 Replies
TimP
Honored Contributor III
289 Views
E5-26xx series show more gain vs. clock speed than earlier models, but it's not linear or 100%. Cost tradeoffs are more complicated than that, involving power consumption, space for more machines, etc. The 95w CPUS are often favored where power efficiency is important. e5-2643 is a higher memory bandwidth chip, presumably requiring a more expensive motherboard (and more cooling), than the others you mention, so it seems unlikely that many people would see the same finalists in their selection process. It might give you more performance per core than the others, even with the higher speed, but I don't think simple estimates will help you much. I doubt that any analyses exist which would answer your questions, beyond what is on the site ark.intel.com
0 Kudos
hareson
Beginner
289 Views
Ok, thank you for your reply. I'll keep this in mind. I'm not really power limited as we don't have that many computers. Cheers
0 Kudos
Reply