Do # of cores or GHz/per cores perform better?

The most demanding Projects are only available to a small percentage of very high-end servers.

Moderators: Site Moderators, PandeGroup

Re: Do # of cores or GHz/per cores perform better?

Postby PinHead » Tue Feb 07, 2012 11:27 pm

Slash_2CPU wrote:Read up on the threads regarding Bulldozer-based CPU's(4200 and 6200-series Opterons) and PPD vs the last generation of K10 CPU's(4100/6100-series Opterons). The K10 is currently a better bang for the buck.


The articles and the pricing structures from both AMD and Intel is kinda what prompted the question. Both companies newest offerings increase the core count but drop the GHz per core. For example 16x2.1 cpu and 2.6x8 at one etailer runs the same $589. So here is where my "all things being equals" falls apart. I can have 16 @2.6GHz and the board will be full or I can have 32@2.1GHz for the same $1,xxx. Or I can have 16 at 2.1GHz for $1,xxx - $589 and drop the second one in when the price drops some.

The Xeon gets a little more tricky when trying to get to 16 because of the models core count and chips per mother board. Their current offering is 1, 2 or 8 cpu per motherboard.
PinHead
 
Posts: 351
Joined: Tue Jan 24, 2012 3:43 am

Re: Do # of cores or GHz/per cores perform better?

Postby PinHead » Wed Feb 08, 2012 2:13 am

bruce wrote:
They both come out about the same.

On a machine that runs nothing but FAH, that's not an issue, but you'd have to ask yourself whether in the actual environment you're going to be running, would the "typical" differences between the slowest and fastest thread be any different depending on whether you had 16 threads or 32 threads. I'm not even sure how to predict an answer to that question.


So on the slowest thread comment that I see so much of, does that mean that there is "forced feedback" code into other threads ( slowdown gets multiplied ) or does that mean that other threads finish their work fast but then have to wait ( NOP ) until the other threads finish?

Sorry, I know I ask PIA questions; but there is a method to my madness! :shock:
PinHead
 
Posts: 351
Joined: Tue Jan 24, 2012 3:43 am

Re: Do # of cores or GHz/per cores perform better?

Postby codysluder » Fri Feb 10, 2012 2:58 am

Not PIA questions if uou don't know the answer.

Mostly I think it's a NOOP answer. If you compute the positions of all the atoms and half (or any other fraction) are finished first and then You compute the forces between every pair atoms, how much error will be wiil be introduced into the the forces if you use the position for some of the atoms from a previous step?.
codysluder
 
Posts: 2128
Joined: Sun Dec 02, 2007 12:43 pm

Re: Do # of cores or GHz/per cores perform better?

Postby PinHead » Fri Feb 10, 2012 4:05 am

:) Well even though I understand what you are saying about margin of error in a math calculation, I'm just saying that something that moves that fast isn't waiting on a step to complete before it starts the next one. Just like you don't wait on your left foot to stop before you start to move your right foot when walking. Well, unless your stalking a deer. :D

Either way, just trying to figure out how SMP thinks, so thanks for the info and I appreciate it!
PinHead
 
Posts: 351
Joined: Tue Jan 24, 2012 3:43 am

Re: Do # of cores or GHz/per cores perform better?

Postby Jesse_V » Fri Feb 10, 2012 4:11 am

PinHead wrote:Just like you don't wait on your left foot to stop before you start to move your right foot when walking. Well, unless your stalking a deer. :D


Congratulations. I believe you just won the prize for one of the most awesome posts in the last month or so. That's worth at least 50k points. :D
Pen tester at Cigital/Synopsys
User avatar
Jesse_V
 
Posts: 2773
Joined: Mon Jul 18, 2011 4:44 am
Location: USA

Re: Do # of cores or GHz/per cores perform better?

Postby Grandpa_01 » Fri Feb 10, 2012 4:12 am

from what I understand smp is stalking that deer :lol:
Image
2 - SM H8QGi-F AMD 6xxx=112 cores @ 3.2 & 3.9Ghz
5 - SM X9QRI-f+ Intel 4650 = 320 cores @ 3.15Ghz
2 - I7 980X 4.4Ghz 2-GTX680
1 - 2700k 4.4Ghz GTX680
Total = 464 cores folding
User avatar
Grandpa_01
 
Posts: 1757
Joined: Wed Mar 04, 2009 7:36 am

Re: Do # of cores or GHz/per cores perform better?

Postby PinHead » Sun Feb 12, 2012 3:58 pm

Well, it definately looks like I need more of one or the other!

Trying my 1st bigadv WU ( 6904 ) on the 6272 and FahMon says it will miss the prefered deadline by about 13h 30 minutes.

Code: Select all
[11:19:33] Project: 6904 (Run 0, Clone 19, Gen 50)
[11:19:33]
[11:19:33] Assembly optimizations on if available.
[11:19:33] Entering M.D.
                         :-)  G  R  O  M  A  C  S  (-:

                   Groningen Machine for Chemical Simulation

                            :-)  VERSION 4.5.3  (-:

        Written by Emile Apol, Rossen Apostolov, Herman J.C. Berendsen,
      Aldert van Buuren, Pär Bjelkmar, Rudi van Drunen, Anton Feenstra,
        Gerrit Groenhof, Peter Kasson, Per Larsson, Pieter Meulenhoff,
           Teemu Murtola, Szilard Pall, Sander Pronk, Roland Schulz,
                Michael Shirts, Alfons Sijbers, Peter Tieleman,

               Berk Hess, David van der Spoel, and Erik Lindahl.

       Copyright (c) 1991-2000, University of Groningen, The Netherlands.
            Copyright (c) 2001-2010, The GROMACS development team at
        Uppsala University & The Royal Institute of Technology, Sweden.
            check out http://www.gromacs.org for more information.


                               :-)  Gromacs  (-:

Reading file work/wudata_07.tpr, VERSION 4.5.4-dev-20110530-cc815 (single precision)
[11:19:41] Mapping NT from 16 to 16
Starting 16 threads
Making 1D domain decomposition 16 x 1 x 1
starting mdrun 'Overlay'
12750000 steps,  51000.0 ps (continuing from step 12500000,  50000.0 ps).
[11:19:50] Completed 0 out of 250000 steps  (0%)
[12:17:19] - Autosending finished units... [February 12 12:17:19 UTC]
[12:17:19] Trying to send all finished work units
[12:17:19] + No unsent completed units remaining.
[12:17:19] - Autosend completed
[12:46:55] Completed 2500 out of 250000 steps  (1%)
[14:15:25] Completed 5000 out of 250000 steps  (2%)
[15:44:19] Completed 7500 out of 250000 steps  (3%)


PinHead
 
Posts: 351
Joined: Tue Jan 24, 2012 3:43 am

Re: Do # of cores or GHz/per cores perform better?

Postby Nathan_P » Sun Feb 12, 2012 4:02 pm

PinHead wrote:Well, it definately looks like I need more of one or the other!

Trying my 1st bigadv WU ( 6904 ) on the 6272 and FahMon says it will miss the prefered deadline by about 13h 30 minutes.

Code: Select all
[11:19:33] Project: 6904 (Run 0, Clone 19, Gen 50)
[11:19:33]
[11:19:33] Assembly optimizations on if available.
[11:19:33] Entering M.D.
                         :-)  G  R  O  M  A  C  S  (-:

                   Groningen Machine for Chemical Simulation

                            :-)  VERSION 4.5.3  (-:

        Written by Emile Apol, Rossen Apostolov, Herman J.C. Berendsen,
      Aldert van Buuren, Pär Bjelkmar, Rudi van Drunen, Anton Feenstra,
        Gerrit Groenhof, Peter Kasson, Per Larsson, Pieter Meulenhoff,
           Teemu Murtola, Szilard Pall, Sander Pronk, Roland Schulz,
                Michael Shirts, Alfons Sijbers, Peter Tieleman,

               Berk Hess, David van der Spoel, and Erik Lindahl.

       Copyright (c) 1991-2000, University of Groningen, The Netherlands.
            Copyright (c) 2001-2010, The GROMACS development team at
        Uppsala University & The Royal Institute of Technology, Sweden.
            check out http://www.gromacs.org for more information.


                               :-)  Gromacs  (-:

Reading file work/wudata_07.tpr, VERSION 4.5.4-dev-20110530-cc815 (single precision)
[11:19:41] Mapping NT from 16 to 16
Starting 16 threads
Making 1D domain decomposition 16 x 1 x 1
starting mdrun 'Overlay'
12750000 steps,  51000.0 ps (continuing from step 12500000,  50000.0 ps).
[11:19:50] Completed 0 out of 250000 steps  (0%)
[12:17:19] - Autosending finished units... [February 12 12:17:19 UTC]
[12:17:19] Trying to send all finished work units
[12:17:19] + No unsent completed units remaining.
[12:17:19] - Autosend completed
[12:46:55] Completed 2500 out of 250000 steps  (1%)
[14:15:25] Completed 5000 out of 250000 steps  (2%)
[15:44:19] Completed 7500 out of 250000 steps  (3%)




Add another CPU when you have the funds, in the meantime let it rip through some SMP WU.
Image
Nathan_P
 
Posts: 1442
Joined: Wed Apr 01, 2009 9:22 pm
Location: Jersey, Channel islands

Re: Do # of cores or GHz/per cores perform better?

Postby MtM » Sun Feb 12, 2012 4:20 pm

PinHead wrote::) Well even though I understand what you are saying about margin of error in a math calculation, I'm just saying that something that moves that fast isn't waiting on a step to complete before it starts the next one. Just like you don't wait on your left foot to stop before you start to move your right foot when walking. Well, unless your stalking a deer. :D

Either way, just trying to figure out how SMP thinks, so thanks for the info and I appreciate it!


You can't start moving the other foot if you still have your weight on it. You need to complete the previous step first and shift your load to that foot to be able to begin the next step. :eugeek:
MtM
 
Posts: 3054
Joined: Fri Jun 27, 2008 2:20 pm
Location: The Netherlands

Re: Do # of cores or GHz/per cores perform better?

Postby Joe_H » Sun Feb 12, 2012 6:07 pm

Since the 6272 just has 8 FPU's for those 16 threads, it basically is only just a bit better than a 8 core CPU for processing F@H. With high enough clock speed the design could do the bigadv WU's within the deadline, but as mentioned you would be better off with plain SMP until you can add an additional 6272.
Image

iMac 2.8 i7 12 GB smp8, Mac Pro 2.8 quad 12 GB smp6
MacBook Pro 2.9 i7 8 GB smp3
Joe_H
Site Admin
 
Posts: 4533
Joined: Tue Apr 21, 2009 4:41 pm
Location: W. MA

Re: Do # of cores or GHz/per cores perform better?

Postby SASinUtah » Wed Feb 15, 2012 3:05 am

I just replaced my p650 AMD dual-core (2.6 Mhz) with a p960 quad (1.8 Mhz) in my laptop; no other hardware changes of any kind were made. The only relevant difference I can find between the two CPUs (according to CPU-World.com) is that the p650 has a 2x1 Mb L2 cache, where the p960 has a 4x512 Kb L2 cache; both have the same HT speed, same SSE features, same size L1 caches per core, same TDP. After 16 steps of Run 10128 (R96, C1, G42) with nothing else running on the system, the p650 had an average TPF of 39:00 and a PPD of 1365; the p960, after 5 steps of the same WU, had a TPF average of 27:30 and a PPD of 2210.

Based upon my small experiment, I would hazard that more cores, even running at a slower clock speed, is more effective for F@H SMP than fewer cores at a higher clock speed.
SASinUtah
 
Posts: 15
Joined: Thu Jan 19, 2012 3:08 am
Location: Middle of Nowhere, Utah

Re: Do # of cores or GHz/per cores perform better?

Postby k1wi » Wed Feb 15, 2012 3:30 am

GHz,
(p650) 2 * 2.6GHz = 5.2GHz,
(p960) 4 * 1.8GHz = 7.2GHz (38% higher)

TPF
(p650) 39:00
(p960) 27:30 (42% lower)

PPD
(p650) 1365
(p960) 2210 (61% higher)

The only thing lacking from the analysis is price - it is really hard to find the price of a mobile CPU that is likely no longer sold individually!

My analysis would be that [for these two CPUs] the difference in performance, once you account for the increase in clock speed, is marginal. Knowing the relative price of the CPUs would be beneficial. The QRB skews the results in terms of PPD as it amplifies the difference in performance.
k1wi
 
Posts: 1152
Joined: Tue Sep 22, 2009 10:48 pm

Re: Do # of cores or GHz/per cores perform better?

Postby bruce » Wed Feb 15, 2012 3:49 am

Same calculations ... I just got distracted before I posted them:

2.6 GHz x 2 = 5.2
1.8 GHz x 4 = 7.2

7.2 > 5.2 so you have upgraded your system by a factor of 1.38 provided SMP scales perfectly.

39:00 / 27:30 = 2340/1650 = 1.42 so SMP scales slightly better than expected, provided your TPF measurements are very accurate, but given the way you estimated those values, they're probably room for that 3% error. I'm not sure we can really conclude anything concrete, but you're welcome to gather some additional data or report a figure of uncertainty associated with the TPFs and we'll see where we go from there.
bruce
 
Posts: 22616
Joined: Thu Nov 29, 2007 10:13 pm
Location: So. Cal.

Re: Do # of cores or GHz/per cores perform better?

Postby PinHead » Wed Feb 15, 2012 4:32 am

The percent increase in overall GHz calc shows 38.46%, but the overall time decrease calc shows 29.4%. Showing about 9% difference, given a 50% increase in cores at a lower clock rate. I am not familiar with the 650 and 960 to know if they are of the same structure and pipeline length; but you have given me a starting point.

Thank You

off to the AMD website
PinHead
 
Posts: 351
Joined: Tue Jan 24, 2012 3:43 am

Re: Do # of cores or GHz/per cores perform better?

Postby SASinUtah » Wed Feb 15, 2012 5:04 am

Kiwi - I paid $70 for the used p960; the p650 came with the system (which I purchased refurbished from HP), so the price differential would be difficult to estimate. I purchased it primarily for folding, since the laptop is on 24/7. At 25 watts TDP, I did not think I could go wrong, especially now that the voltage reported in CPUID was 1.125v for the p650 and 0.987v for the p960, possibly reducing the power usage even more relative to output.

bruce - the TPF were averages over the 16 and 5 steps, respective, rounded to the nearest :10, so there is room for error. Of what type of additional data would be helpful?
SASinUtah
 
Posts: 15
Joined: Thu Jan 19, 2012 3:08 am
Location: Middle of Nowhere, Utah

PreviousNext

Return to SMP with bigadv

Who is online

Users browsing this forum: No registered users and 2 guests

cron