Main > Main Forum

4Ghz Core 2 Duo vs M.A.M.E. 0.120 (benchmark results)

<< < (26/104) > >>

TheManuel:

--- Quote ---How does this effect MAME if you drop from dual channel to single channel, well very little at 500mhz (2000 FSB) DDR2-1000 with a E6850 with 4mb of cache like I'm running, about 2-3% across the board in fact as you can see in the benchmarks below.

The catch is that while MAME is basically CPU limited and doesn't appear to require huge memory bandwidth, these result are on my system and thus not the norm, the effect of dropping to only 6.4Gb/s memory bandwidth with a single DDR2-800 Dimm on say a E4*00 series with only 2mb CPU Cache will probably be much more noticeable that the effect of dropping to 8.0GB/s with 4mb of CPU cache on my system.
--- End quote ---

tar-nz:
I think you are well intentioned and you've got the theory down in great shape for the mid-terms :-)
However, I think in practice, it can all be mis-leading.  When I tried dual vs single channel on my setup, just out of curiosity, I did not see any significant difference.  I did not run the benmchmarks in a very disciplined manner but there was nothing that jumped out at me.  I will try to run some tomorrow night (I have to do Tiger-Heli's benchmarks tonight :-) ).  Beside, the theoretical bandwith you quote above is just that, theoretical.  Please run some synthetic benchmarks on your system using Sandra or some other software and report back on your bandwidth.  I'm sure you will notice that it falls way short of theoretical.  In fact, for a more eye-opening experiment, run this bechmark on single channel mode as well and see if dual channel gets anywhere close to twice the bandwidh of dual channel.  Oh, and sysnthetic benchmarks are always the most optimistic test you can run and real world performance is usually lower anyway.
Overclocking webistes tend to get too hung-up on memory bandwith but every time I test different setups I don't see any differences worth writing home about so I'd hate for people planning out their systems to spend too much time and money for extra bandwidth when it does not do much in practice.  Looking at your benchmarks, most games improved by less than 2% (dividing dual channel percentage by single channel percentage), some even reduced performance while a few improved by as much as 6% and this is in the extreme case of single vs dual channel.  So in the end you have to ask yourself if 6% in a few isolated games is worth going after. 

Nevertheless, going dual channel probably makes financial sense since most modern motherboards support it and a pair of matched memory sticks is only a few bucks more than a single stick with the same total memory.  Even if you are not sure it's going to make much difference, you just protect yourself for the unknown for a few bucks more.  However, spending a lot of money on super fast memory doesn't make much sense to me.

I still owe you some benchmarks to put some evidence behind my claims.

Tiger-Heli:
Taz did post both specs and the info was useful - and he said his system was likely less affected by it than most.

I still want to see TheManuel's benchmarks, but I did get a chance to test three of the game's out here on the work, ermmm, test system and wanted to post them here:

This is an Hewlett Packard E6400 running at stock clock of 2.13Ghz, WinXP, 2G of Ram (dual channel).  I used standard (non multi-core optimised MAME 0.121) with total default settings and didn't run anything specific, just looked at F11 (I also had Access, Outlook, and Seamonkey, and maybe Word running in the background):

Crusnusa ran at pretty near 100%, dropped to 90% at times.
Swtrilgy ran at around 34% from what I could tell
Sfrush ran at about 53-57% but I just got past the startup screens and closed it out (won't really work without a joystick).

Not bad at all and looks like if I get to about 3.0 Ghz, those games are likely playable.

I've about decided how I want to tackle this is a stock E2xxx processor and cooler, either a P35 or 965 mobo (probably Gigabyte or Abit, maybe Asus, Asrock, or BioStar (are these okay?) (also on the 965, I assume I want P965 rather than G or Q since I won't have integrated graphics, but are there other disadvantages to a G965 or Q965?) also I saw FoxConn uses solid capacitors also - any opinions on them, one stick of 1G ram, and a cheap vid card (might need a new power supply also - recommmended brands? - I've used Fortron in the past).

I'll probably see how this does stock, might try to overclock and hit close to 3Ghz.  May add a second stick of 1G ram and a better CPU cooler at that time.

Like TheManuel said, I'm really not too concerned if CrusnUSA runs at 91% instead of a possible 93%, but getting it to 91% from about 39% (Last I remember - XP Barton 2800+) is important.

TheManuel:
The info was useful indeed.
Taz is the man.  I certainly was not meaning to diminish his comments in any way.
He obviously knows this stuff better than I do.

taz-nz:

--- Quote from: TheManuel on November 29, 2007, 06:54:56 pm ---The info was useful indeed.
Taz is the man.  I certainly was not meaning to diminish his comments in any way.
He obviously knows this stuff better than I do.


--- End quote ---

Ha I get where your coming from, there is way  to much hype around bandwidth and getting that last 100 point in this benchmark or that. I probably shouldn't have posted the Theorical numbers knowing good well it's impossible to reach those figures in the real world, in fact the best Intel memory controllers are only about 50% efficient, the intergated memory controller in the 939 Ahtlon64 on the other hand is 90%+ efficient and get much closer to the theorical number than any other system setup.

Memory and system bandwidth play a greater of lesser roll in preformance depend on your the software your running, But there is point where if you decrease the available bandwidth to the CPU enough no amount of Cache memory will stop a noticably drop in prefromance, now it appears we're lucky with MAME and it's memory bandwidth requirements are fairly low, but a drop in memory bandwidth will still effect it's preformance to a greater or lesser degree depend on your over all system setup.

I lot of my knowledge of effects of bandwidth on system preformance come from my P4 days, when I ran a 2.8ghz P4 northwood (512k cache) @ 3.5ghz 250mhz FSB with DDR-500 , that system was faster than a 3.6ghz P4 Prescott (2mb cache) with DDR-400, it was almost as fast as a 3.8ghz P4 Prescott. The extra cache memory in Prescott Core could not make up for the lack of system & memory bandwidth, the P4 Core simply needed more bandwidth to keep it work hard at high clock speeds.  I also spend a lot time disigning Dual CPU Quad Core Xeon server setup for customers at work, these use a Quad Channel memory setup to supply enough memory bandwidth to keep 8 cores working hard. 

Now not all of this knowledge directly applies to Core 2 Duo, but the basic rules still apply bandwidth matters it just depend of your setup how much. I'm just trying to get people to look at their system setups as a whole rather than just focusing on the CPU clock speed, the best preformance always comes from balanced systems.

I hope you get where I'm trying to come from.



--- Quote from: Tiger-Heli on November 29, 2007, 12:55:56 pm ---I've about decided how I want to tackle this is a stock E2xxx processor and cooler, either a P35 or 965 mobo (probably Gigabyte or Abit, maybe Asus, Asrock, or BioStar (are these okay?) (also on the 965, I assume I want P965 rather than G or Q since I won't have integrated graphics, but are there other disadvantages to a G965 or Q965?) also I saw FoxConn uses solid capacitors also - any opinions on them, one stick of 1G ram, and a cheap vid card (might need a new power supply also - recommmended brands? - I've used Fortron in the past).

--- End quote ---

Watch yourself with the 965 chipset boards, there was much larger spread in overclockablity on these boards, the P35 chipset handle high FSBs better and the P35 boards tend to designed to handle those high FSBs better, there are some good 965 overclocking boards out there but you'll probably find you can pick up a P35 chipset board for about the same price. As for brand stick with Gigabyte of ASUS, I was a big fan of Abit up until about 2 years ago, and then they just seamed to loss it, Asrock & Biostar both make cheap rubbish stay away from them. Power supply Cooler master makes some ok 460watt PSU at a good price, but you may not need anything that big depending on what other hardware you plan to run.




 

TheManuel:
A quick post here because I have to run off to a friend's house but did not want to leave Tiger-Heli hanging.
I only managed to run half of the benchmarks:

      1.8GHz      2.7GHz
sfrush      60.20%      84.33%
crusnusa                   99.21%      150.16%
swtrilgy                   25.64%      40.82%
ridgerac                   44.81%      66.17%

Some games scale linearly, some don't.
I ran taz's command line:
mamep.exe -noautoframeskip -frameskip 0 -seconds_to_run 240 -nothrottle -nosleep -video ddraw -skip_gameinfo -effect none -nowaitvsync -noreadconfig -mt -rompath f:\games\mame\roms sfrush

For sfrush, I press P1 button at the calibration screen to get rid of it in both cases.
I did not run swtrilgy to completion for the 1.8GHz (only 187sec) because I am in hurry so it may be somewhat biased.

Tomorrow I will run the other half of the benchmarks and also check the in-game FPS while actually playing the games with the CPU at 3.0GHz since this is ultimately what matters.

Regards.

Taz:
I got a glimpse of your message but will read it thoroughly tomorrow.

Navigation

[0] Message Index

[#] Next page

[*] Previous page

Go to full version