CFD Online Discussion Forums

CFD Online Discussion Forums (http://www.cfd-online.com/Forums/)
-   Hardware (http://www.cfd-online.com/Forums/hardware/)
-   -   RAM speed, is it worth the fastest? (http://www.cfd-online.com/Forums/hardware/95432-ram-speed-worth-fastest.html)

Jordi December 17, 2011 09:14

RAM speed, is it worth the fastest?
 
Has anybody experience with RAM speeds? Does a 1800 MHz RAM work (18/16) faster than a 1600 MHz RAM? (I mean speed in CFD computing time). Do you think it is worth the extra cost?

kyle December 17, 2011 13:42

If you look at the STARS Euler3d benchmark on this page, you will see that the difference between 1333mhz memory and 1600mhz memory speeds up the simulation by 10%. 1333mhz to 2133mhz speeds it up by over 20%.

In contrast I ran benchmarks on my cluster at a stock CPU clock speed of 3.4ghz and overclocked to 4.0ghz. The speedup was only ~3%.

Intel does not allow you to run memory faster than 1600mhz on server class hardware, so consumer hardware is potentially faster than server class hardware. The new i7 3960X was born for CFD. You can run 2400mhz memory on four channels. That is a lot of bandwidth.

Jordi December 17, 2011 17:45

overclocking never speeds anything... the consumer is so mesmerized about the MHz that cannot realize the bottleneck of computing is currently in data transfer not in processing.

This website has lots of good info, thanks. But CPUs are tested with different motherboards (of course), memory size and speed, so the results are somehow subjective, I think.

kyle December 17, 2011 18:41

Sorry I linked to the wrong article there. Here is the one that benchmarks the same CPU and motherboard with varying memory speed:

http://techreport.com/articles.x/20377/2

Jordi December 18, 2011 04:17

It seems that the difference between 1600 and 2133 MHZ SDRAM is some 10% in speed. I think this does not justify to me the extra expense, not only the RAM cost but also the motherboard. Thanks a lot, you provided me the very right information!!

evcelica March 21, 2012 21:25

This comparison was done with sandy bridge which has dual channel memory. I wonder what the result would be with triple or quadruple channel memory since it also increases the banwidth. I would guess it would make even less of a difference since other parts of the system would become the bottleneck before memory bandwidth. can anyone comment?

rmh26 March 22, 2012 19:44

The answer would really depend on the software and the problem size. For a small problems the data can be handled efficiently by the cache but as the number of variables grows or the dimensionality of the problem increases the data won't fit as well in the cache and the cpu will have to go to the ram itself more often. This is where faster ram will help. But the problem size at which it becomes memory bound isn't really know until you run the code.

You could try under-clocking your current ram and running simulations to see the change in speed and if it is worth buying faster ram.


All times are GMT -4. The time now is 17:27.