mersenneforum.org  

Go Back   mersenneforum.org > Great Internet Mersenne Prime Search > Hardware

Reply
 
Thread Tools
Old 2016-03-29, 13:48   #1
Fred
 
Fred's Avatar
 
"Ron"
Jan 2016
Fitchburg, MA

97 Posts
Default Cas Latency

Recognizing that memory speed/bandwidth is king, and all other factors being even, is it fair to assume that cas latency has a fairly significant impact on results (LL testing)?

In particular, I'm looking at cas ratings of 12 vs 15 for an extra $16 :

http://www.newegg.com/Product/Produc...D=3938566&SID=

vs

http://www.newegg.com/Product/Produc...D=3938566&SID=
Fred is offline   Reply With Quote
Old 2016-03-29, 14:32   #2
Mark Rose
 
Mark Rose's Avatar
 
"/X\(‘-‘)/X\"
Jan 2013

29·101 Posts
Default

I don't recall anyone doing a CAS latency benchmark comparison.

Perhaps underclock your existing RAM and see what effect it has?
Mark Rose is offline   Reply With Quote
Old 2016-03-29, 17:27   #3
Madpoo
Serpentine Vermin Jar
 
Madpoo's Avatar
 
Jul 2014

1100110110102 Posts
Default

Quote:
Originally Posted by Fred View Post
Recognizing that memory speed/bandwidth is king, and all other factors being even, is it fair to assume that cas latency has a fairly significant impact on results (LL testing)?
The CAS latency does play a part in other memory benchmarks so I can guess it would also impact P95 performance in the same way. But as Mark said, I don't know that anyone has done a side-by-side comparison to confirm that hypothesis.

In general though, lower latencies are going to give you better performance overall even at the same memory speed. If it were me, I'd pony up the extra $16 just to be sure.
Madpoo is offline   Reply With Quote
Old 2016-03-29, 17:33   #4
science_man_88
 
science_man_88's Avatar
 
"Forget I exist"
Jul 2009
Dumbassville

26·131 Posts
Default

Quote:
Originally Posted by Madpoo View Post
If it were me, I'd pony up the extra $16 just to be sure.
I'm not into hardware or software all that much but you'd pay ~47% more to have 25% more speed ? as that's what I would get thinking of rough math in my head. it's also only because they are both SDRAM that they are simply comparable as DRAM itself would have nano second timings where as SDRAM has clock cycles.

Last fiddled with by science_man_88 on 2016-03-29 at 17:45
science_man_88 is offline   Reply With Quote
Old 2016-03-29, 17:57   #5
mackerel
 
mackerel's Avatar
 
Feb 2016
UK

13·31 Posts
Default

Is the mobo that it will be used in not supporting faster ram? I would think that would have more of a benefit.

I have done a lot of testing in the past, but ram timings wasn't a part of it. If we take the choice as presented, I'll see if I can get one of my systems into a state to test that, but it might take a few hours as I have other things to do first.
mackerel is offline   Reply With Quote
Old 2016-03-29, 18:06   #6
Mark Rose
 
Mark Rose's Avatar
 
"/X\(‘-‘)/X\"
Jan 2013

1011011100012 Posts
Default

Quote:
Originally Posted by mackerel View Post
Is the mobo that it will be used in not supporting faster ram? I would think that would have more of a benefit.

I have done a lot of testing in the past, but ram timings wasn't a part of it. If we take the choice as presented, I'll see if I can get one of my systems into a state to test that, but it might take a few hours as I have other things to do first.
I truly appreciate all the benchmarking you do.

Perhaps we should start an ideal hardware thread that can be stickied for future reference.
Mark Rose is offline   Reply With Quote
Old 2016-03-29, 19:23   #7
tServo
 
tServo's Avatar
 
"Marv"
May 2009
near the Tannhäuser Gate

607 Posts
Default

Quote:
Originally Posted by Mark Rose View Post
I don't recall anyone doing a CAS latency benchmark comparison.
Perhaps no one here using Prime95 but if you google
cas latency anandtech or cas latency tom's hardware

you will see lots of articles about memory tests/comparisons, particularly on the
Anandtech site. They also compare DDR3 vs DDR4, raw speed vs cad latency, etc.
Anandtech's articles are quite detailed and lengthy.

Most of the posts on Tom's Hardware are from the forums of people asking questions.
tServo is offline   Reply With Quote
Old 2016-03-29, 19:36   #8
mackerel
 
mackerel's Avatar
 
Feb 2016
UK

6238 Posts
Default

2400-15-15-15-35
Timings for 4096K FFT length (1 cpu, 1 worker): 15.41 ms. Throughput: 64.88 iter/sec.
Timings for 4096K FFT length (2 cpus, 2 workers): 16.28, 16.26 ms. Throughput: 122.93 iter/sec.
Timings for 4096K FFT length (3 cpus, 3 workers): 20.01, 19.98, 19.92 ms. Throughput: 150.22 iter/sec.
Timings for 4096K FFT length (4 cpus, 4 workers): 25.71, 25.76, 26.57, 26.39 ms. Throughput: 153.24 iter/sec.

2400-12-14-14-35
Timings for 4096K FFT length (1 cpu, 1 worker): 15.41 ms. Throughput: 64.88 iter/sec.
Timings for 4096K FFT length (2 cpus, 2 workers): 16.22, 16.25 ms. Throughput: 123.16 iter/sec.
Timings for 4096K FFT length (3 cpus, 3 workers): 19.75, 19.78, 19.86 ms. Throughput: 151.55 iter/sec.
Timings for 4096K FFT length (4 cpus, 4 workers): 25.85, 25.95, 25.88, 25.66 ms. Throughput: 154.82 iter/sec.

3000-15-17-17-35
Timings for 4096K FFT length (1 cpu, 1 worker): 15.29 ms. Throughput: 65.42 iter/sec.
Timings for 4096K FFT length (2 cpus, 2 workers): 15.66, 15.65 ms. Throughput: 127.75 iter/sec.
Timings for 4096K FFT length (3 cpus, 3 workers): 17.42, 17.44, 17.36 ms. Throughput: 172.35 iter/sec.
Timings for 4096K FFT length (4 cpus, 4 workers): 21.47, 21.38, 21.41, 21.44 ms. Throughput: 186.72 iter/sec.

Test system has an i5-6600k fixed at 3.5 GHz. Ram is dual channel, single rank.

Comparing the two tests at 2400, the tighter timings gives you under 1%. Turn the wick up to 3000 for a more visible boost. 25% more ram clock for 20% more performance. Think we might be a bit ram bandwidth limited here?
mackerel is offline   Reply With Quote
Old 2016-03-29, 19:53   #9
Mark Rose
 
Mark Rose's Avatar
 
"/X\(‘-‘)/X\"
Jan 2013

29·101 Posts
Default

So lower CAS latency helps a little for Prime95, but bandwidth is king.
Mark Rose is offline   Reply With Quote
Old 2016-03-29, 20:03   #10
mackerel
 
mackerel's Avatar
 
Feb 2016
UK

13·31 Posts
Default

Up to a point. On the weekend I was using the same system to see if I could overclock the ram further. I reached 3600 at some slack timings, and it only increased throughput 3.3% over 3000. So it seems 3000 gives you most of the bandwidth you need for this system. Of course, if the CPU clock is faster, you would need faster ram again.

I think this also showed a potential error in my earlier estimates in how rank influences things. I suspected my results were slightly over-estimating the ram bandwidth required. I didn't allow for the change in throughput being non-linear depending on the balance between ram and cpu. It only approaches linear if you're really limited by one or the other, so I'll have to retest that and see what it does.
mackerel is offline   Reply With Quote
Old 2016-03-29, 21:13   #11
kladner
 
kladner's Avatar
 
"Kieren"
Jul 2011
In My Own Galaxy!

22·2,539 Posts
Default

This is an interesting and educational thread.
kladner is offline   Reply With Quote
Reply

Thread Tools


Similar Threads
Thread Thread Starter Forum Replies Last Post
Benchmarks varying FSB, Memory Latency and multiplier S485122 Software 0 2006-11-08 20:21
Help: trying to determine latency on movaps instructions on AthlonXP LoKI.GuZ Hardware 1 2004-01-26 20:05

All times are UTC. The time now is 20:48.

Sat Feb 27 20:48:15 UTC 2021 up 86 days, 16:59, 0 users, load averages: 0.98, 1.46, 1.58

Powered by vBulletin® Version 3.8.11
Copyright ©2000 - 2021, Jelsoft Enterprises Ltd.

This forum has received and complied with 0 (zero) government requests for information.

Permission is granted to copy, distribute and/or modify this document under the terms of the GNU Free Documentation License, Version 1.2 or any later version published by the Free Software Foundation.
A copy of the license is included in the FAQ.