Bursting the Athlon 64 Memory bandwidth bubble

Memory by Gamer @ 2005-05-23

Do you need high speed memory to have the fastest A64 on the block? We don?t think so, we think having memory faster then PC3200 is overkill and does not improve real world performance one bit. We put our money where our mouth is and prove that Gamers don't need expensive memory to enjoy fast Athlon 64 performance.

  • next

System Setup & Benchmarks

So you bought yourself a new Athlon 64 CPU and want to pair it up with high quality components to extract maximum performance from your new PC? What memory should you buy?

It’s safe to say that PC2700 and older memory will not live up to the A64 challenge, starting from PC3200 your system will have all the memory bandwidth it needs to be blistering fast. Back in the glory days of the Athlon XP and Pentium 4 people pushed their FSB upwards to increase the memory bandwidth and this paid off quite well with an increase in performance of ~15% without the CPU being clocked higher!. Companies saw the need for memory rated higher than PC3200 and so the unofficial PC3500/3700/4000/4400/xxxx standards were born.

When Athlon 64 was released to the public many tried their tested tactics on this new platform, getting the HTT (A64’s equivalent to the FSB) as high as possible to gain extra performance. But A64 has an on-die memory controller,unlike the Athlon XP or P4, this controller allows you to run memory asynchronous to the HTT. Running memory async to the system bus is of course nothing new, but what is new, is the fact that performance between a system running memory async versus one running memory synchronous is virtually non-existent.

Sure there will be applications and benchmarks which will show you a difference favoring the synchronous 1:1 running system, but overall this increase in performance is nothing compared to the gains we so accustomed to have when doing the same on Athlon XP or P4.

So here today we are focusing on proving that Gamers don’t need expensive memory to get the most out of their Athlon 64 system.


System Setup

Here’s our test hardware, an Athlon FX CPU has the advantage of being able to increase/decrease the multiplier which helps us in our tests, but it is not needed, the FSB:MEM divider in the BIOS allows you to keep the memory running within specs

  • AMD Athlon A64 FX-55 (13x200)
  • Swiftech H20-120-FB Water Cooling Kit
  • MSI K8N Neo2
  • Sapphire X800XT-PE
  • Samsung SATA 7200 120GB -8 mb (OS)
  • 2X512 MB Gskill PC4400 LE TCCD

    The PC4400 G.Skill memory is also rated CL2 2-2-5 at PC3200, this will come in handy in this test.


    CPU Speeds

    Our first configuration has the CPU running at 2800Mhz with a 14x multiplier, so our memory is running 200Mhz

    Madshrimps (c)



    Next up we decrease the multiplier to 10x and increase the HTT to 280Mhz

    Madshrimps (c)



    Memory Speeds

    If you buy a no-name or budget PC3200 pair of memory sticks, then those will most likely run at CL2.5 3-3-x at 200Mhz. So we configured our G.Skill like this:

    Madshrimps (c)


    When you spend a few more $/€ you can get “eXtreme” sticks which will operate at very tight timings at PC3200 speeds; CL2 2-2-5

    Madshrimps (c)


    Then we have the “Extreme Performance” categorie which will run at insane speeds close to/and over 300Mhz at more relaxed timings. CL2.5 3-3-7 at 280Mhz!

    Madshrimps (c)


    Recently a new series of “Extreme Performance” sticks was launched, most likely equipped with Winbond UTT memory chips, they keep running tight timings all the way up to PC4000 speeds, but require massive amounts of DDR voltage to do so. We would have loved to include a pair of those in our test as we have an OCZ DDR Booster lying ready here, but we don’t have this type of modules in our lab right now.


    Benchmarks

    And this is what you’ve all been waiting for, the hard proof. We start of with a synthetic benchmark. Sisoft Sandra’s memory benchmark measures the maximum bandwidth of your system and reports results back in Mb/s

    Madshrimps (c)


    As you can see, with the memory running synchronous at 280Mhz the bandwidth increases with a quite impressive 1600Mb/s. The difference at PC3200 speeds with different timings is less pronounced: ~60Mb/s

    But do we “play” Sisoft Sandra, SuperPi, 3DMark2001SE and other benchmarks? No!

    Some do “play around with” these benchmarks to obtain the highest scores and I’m sure they’re having a great time doing so.

    But you’re not going to buy more expensive memory just to “show off” are you? If you answered yes to this question, then you should not be reading this article anyway as you already know the outcome.

    For those who do like to get the most bang for their buck, we present you our game benchmarks ->
    • next
    Comment from BAMBI @ 2005/05/23
    WOW i would have expected a much bigger difference!

    would be interesting to see if there would be differences in other apps like encoding, and also have a pp4 comparison.

    Great job on the test though
    Comment from Bosw8er @ 2005/05/23
    article

    Is the human eye capable of noticing the difference between 85 or 93 FPS ?
    Comment from jmke @ 2005/05/23
    24 FPS is actually enough to trick the eye for movies to be "smooth"; PC games however sit more comfortable at 60FPS; and First Person Shooters can benefit from >100FPS, although this depends from person to person.

    anyway, noticing the difference between anything >60FPS is imho very hard, the advantage of having a higher average FPS is that you also get higher frame rates when there is a lot of action on screen and there is a dip in framerate
    Comment from Jaco @ 2005/05/23


    low latency (2-2-2 timings) is the magic word for A64...
    you don't need all that bandwidth
    Comment from jmke @ 2005/05/23
    and even that is disputable; CL2.5 3-3-7 seems to be holding up quite good also
    Comment from kristos @ 2005/05/23
    very nice article

    someone also compared the effects of cas and ras to cas on XS a few months ago and his results were comparable.

    http://www.xtremesystems.org/forums/...ad.php?t=48634
    Comment from perdomot @ 2005/05/23
    I would also like to see how bandwidth affect encoding with popular software like DVD Shrink and TMPEG encoder. Any chance someone can try this?
    Comment from Sidney @ 2005/05/23
    I am now able to run the GSkill @ 1.5-2-2-5 with 3 Volt 200 Mhz; hardly any difference with the exception of dropping Sandra mem bandwidth on benchmark; otherwise, there is no performance gain to speak of.

    Still bandwidth >6200mb/s; too bad my DVD Shrink is in the Prescott system, otherwise, I will burn a copy DVD disks to test the speed ..... then again, it also depends on HD.
    Comment from Haut^Karl @ 2005/05/23
    While the premise of the article is interesting, the testing is flawed. 'gamer' set out to prove that increasing memory bandwidth is unnecessary to gamers so he selected two HTTs while holding timings the same. Good start. Then for testing purposes, he/she chooses 2 graphics limited games(Halflife not as much) AND runs them at 10x7 and 12x8 *maximum quality* to add insult to injury! Even with older games(read Quake3) this taxes the graphics card more than the CPU. The article simply reillustrates that those 2 games are graphics limited and abandons the premise of memory/cpu bandwidth.

    'Gamer' could you retest using 6x4, possibly 8x6, using a less graphics limited game with *minimal* graphics settings to truly test memory/cpu bandwidth? I am certainly interested in the findings.

    IMHO, I don't believe we "gamers" only play the most recent games and that's all. As you can see from the responses above, we re-encode DVDs & video and play many, many other games & applications.

    Thanks
    Comment from jmke @ 2005/05/23
    Hello Haut^Karl, welcome to the forums

    what USE does it have to benchmark games at lower resolutions with less detail when you spend money on a system to play in high detail at high resolutions!

    there is no use; surely at 640x480 you will see a larger difference, but that does not prove anything, since you don't play games at that resolution or graphics detail.

    about the older games: with a new system you can run older games at higher resolutions with more detail, FSAA/AS enabled and more, again, enjoying the fact that you bought a fast A64 system. there is no need to drop your details just so can say "look my expensive memory runs this game faster at 640x480".

    the whole point of the article is to proof that Gamers don't need expensive memory to make their system run faster when playing games

    the same point was made in a previous article here when it comes down to CPU power; a 3ghz A64 is too powerful for ANY gamer, as the GPU is limiting factor here. http://www.madshrimps.be/gotoartik.php?articID=231

    about the DVD encoding: we are working on it, I'm interesting in the results as much as you
    Comment from Haut^Karl @ 2005/05/24
    Hi. Glad to finally be a member in these forums!

    The reason you employ lower resolutions is to test the changes in memory bandwidth. More bandwidth means you can move many more frames per second than with less bandwidth. Therefore we would want to provide as many frames as possible in order to fill a fatter pipeline, so to speak. This also removes the GPU as the limiting factor since low res, simple graphics are easy for it to do. This is how you would test the bandwidth issue.

    "the whole point of the article is to proof that Gamers don't need expensive memory to make their system run faster when playing [Doom3 and Halflife2]" would be a more correct statement since those 2 games are undeniably graphically intense making ram and CPU speed less relevant. I could agree with that. However if 'gamer' were to choose some CPU limited games like AoE 2, Sims(pick your flavor), or Jedi Knight 2 we could see some relevant results.

    But to choose GPU limited examples to highlight changes in memory bandwidth(or CPU speed) is the flaw in the article and meaningless at best. Sorry gamer

    For those readers who don't follow what we are discussing, Anand has a good illustration here: http://www.anandtech.com/cpuchipsets...oc.aspx?i=2149

    The first image shows at lower resolutions, more CPU/memory bandwidth will allow the GPU to crank out more frames. <-- This is what we should see in this article. At higher resolutions it's the GPU that is holding you back and more CPU Mhz/bandwidth will show little effect. The second image shows the 9800 is the bottleneck as it is clearly a generation behind the 6800 and has only half the pipelines.

    Anywho! The DVD encoding will be relevant but will not concur with the original article. Since encoding is all about CPU speed and memory bandwidth, it will jump and giggle everytime you give it more bandwidth(either through HTT or timings).

    Off to walk the dog now!!
    Comment from Sidney @ 2005/05/24
    Just came back from walking my Lab. Well, I thought the article focuses on where is the best to spend the money; memory with tight timings or better graphic card. For example, a pair of 512 mB PC3200 priced below $100 versus low timings PC4400 costing over $200, the price difference could enable a gamer to reach for better graphic card. Naturally, it is better to have both. Bottom line, where to get the most from your hard earned money.
    Comment from unusualfire @ 2005/05/24
    Wasn't the tests done on windows xp?
    Maybe the more expensive chips will shine in windows xp x64.
    Comment from jmke @ 2005/05/24
    Quote:
    Originally posted by Haut^Karl
    But to choose GPU limited examples to highlight changes in memory bandwidth(or CPU speed) is the flaw in the article and meaningless at best.
    the fact that you don't see the changes in bandwidth is the whole point of the article;

    I can run Jedi Knight 2 in 1280x1024 4xFSAA/8xAS and you'll see no gains in performance again when overclocking the HTT on the A64.

    if you want to play Sims then you don't need an Athlon 64, as that game will run fine on a Pentium I 166mhz

    there is no denying that the Doom3, Quake3 and HL2 engine will be widely used for newer games, and this article points out that your PC will be ready to handle them, regardless of what kind of memory you have installed, as the difference between expensive tight timings / high Mhz and cheap PC3200 stuff is virtually not existant.


    Quote:
    For those readers who don't follow what we are discussing, Anand has a good illustration here: http://www.anandtech.com/cpuchipset...doc.aspx?i=2149

    The first image shows at lower resolutions, more CPU/memory bandwidth will allow the GPU to crank out more frames. <-- This is what we should see in this article.
    On the contrary, that is not what we should see in this article, since you should not be playing Doom 3 at 640x480... and if you are , and you have money to spend, should you
    a) spend that money on more expensive memory so you now run Doom3 10% faster at 640x480, or..
    b) spend that money on a new videocard and enjoy Doom3 at 1024x768 and beyond.

    see my point?


    Quote:
    Originally posted by unusualfire
    Wasn't the tests done on windows xp?
    Maybe the more expensive chips will shine in windows xp x64.
    the difference between games running 32-bit vs 64-bit is.. well.. not there
    http://www.hardocp.com/article.html?art=NzY1LDEx

    The focus of this article is to determine if upgrading to Windows XP Professional x64 Edition with an AMD Athlon64 would give us a better gaming experience in current games. We used eight games for our evaluation including one that supported Win64 AMD64 specifically.

    What we found isn’t too shocking really, but rather reassuring. In all the 32-bit games tested, we saw overall static performance using Windows XP Pro x64 Edition. The only game in our lineup that had a specific Win64 AMD64 instruction path, The Chronicles of Riddick, actually performed worse in 64-bit than it did in 32-bit Windows.
    Comment from perdomot @ 2005/05/24
    JMKE,
    If the extra bandwidth isn't useful for gaming, what is it good for? I am considering getting faster ram for my rig because I am overclocking. Right now, I am running my A64 3000+ at 255 X 9 with a memory divider that brings my value ram down to 163. I was thinking of getting some PC4000 so I could run things at 1:1 right around 250 HTT but if there is no performance increase anywhere, particularly video encoding, I won't spend the money. What would you advise? Thanks.
    Comment from jmke @ 2005/05/24
    hey perdomot; CPU intensive tasks which rely solely (or mostly) on the CPU alone will show benefits; but this increase will not be that impressive either.

    I would advise to spend money on faster HDD, videocard, burner, etc, before buying speedier RAM; or maybe consider buying MORE ram; as having more does not hurt, and you might be able to run a RAMDISK and increase performance tenfold
    Comment from jmke @ 2005/05/24
    solidshot @ futuremark ran some benchmarks, including rendering, 3dmark and superpi, his results concur with what I stated in my post above ^^


    quote: http://discuss.futuremark.com/forum/...=5&o=0&fpart=1

    All tests were run on the following system:
    AMD 64 3200+ Winchester
    DFI NF4 Ultra-D with 5/10-3 Bios
    Sapphire Radeon X850XT @ Stock (no tweaks, no aa/af unless specificed. Drivers are Catalyst 5.5)
    Windows XP SP2
    Every memory option, including volts were the same for all tests, the only difference in timings were the changes from 2-2-2-5 to 2.5-3-3-7.



    >>>>>>>>>>>>>>>>>>>>>>>>>2.6GHZ 260HTT 216FSB 10x260 2-2-2-5 5/6 Divider (166)

    CPU - 2.6GHZ 10x260 (260 HTT)
    MEM - 216fsb 2-2-2-5 1t (166 Divider, DRAM/FSB :5/06)
    X850XT @ Stock
    DFI NF4 Ultra-D 5/10-3 Beta Bios

    SuperPI 1MB: 34.45s
    Everest Latency: 40.6ns
    Everest Bandwidth: 6,400mb/s

    LightWork Render Bench:
    Setting: 800x600
    Results: 5min 25s

    3DMark01 SE:
    Results: 26,827

    Doom 3 timedemo demo1:
    Settings: 1024x768 High Quality, noAA/8xaf (af automatically set to 8x @ high quality)
    Results: 95.2 FPS
    1280x1024 Test (High Quality, same settings)
    Results: 75.0 FPS

    Half-Life 2 CS:Source Stress Test
    Settings: 1024x768. Model High, Texture High, Water Reflect All, Shadow High, AA none, Filtering Trilinear, Shader High, Vsync Disabled
    Results: 144.82 FPS
    1280x1024 Test (same settings)
    Results: 132.95 FPS

    -----------------------------------------------------------------------------------

    >>>>>>>>>>>>>>>>>>>>>>>>>2.6GHZ 260HTT 260FSB 10x260 2.5-3-3-7 1/1 Divider (200)

    CPU - 2.6GHZ 10x260
    Mem: 260FSB 2.5-3-3-7 1t 1:1
    X850XT @ Stock
    DFI NF4 Ultra-D 5/10-3 Beta Bios

    SuperPI 1MB: 32.95s
    Everest Latency: 39.7ns
    Everest Bandwidth: 7,337mb/s

    LightWork Render Bench:
    Setting: 800x600
    Results: 4min 4s

    3DMark01 SE:
    Results: 27,334

    Doom 3 timedemo demo1:
    Settings: 1024x768 High Quality, noAA/8xaf (af automatically set to 8x @ high quality)
    Results: 96.0 FPS
    1280x1024 Test (High Quality, same settings)
    Results: 77.8 FPS

    Half-Life 2 CS:Source Stress Test
    Settings: 1024x768. Model High, Texture High, Water Reflect All, Shadow High, AA none, Filtering Trilinear, Shader High, Vsync Disabled
    Results: 145.71 FPS
    1280x1024 Test (same settings)
    Results: 133.06 FPS


    -------------------

    Breakdown Results for 2-2-2-5 216fsb and 2.5-3-3-7 260fsb while cpu maintained 2.6GHZ

    SuperPI 1MB
    2-2-2-5: 34.45s
    2.5-3-3-7: 32.95s

    Everest Latency
    2-2-2-5: 40.6ns
    2.5-3-3-7: 39.7ns

    Everest Bandwidth
    2-2-2-5: 6,400mb/s
    2.5-3-3-7: 7,337mb/s

    RenderBench 800x600
    2-2-2-5: 5min 25s
    2.5-3-3-7: 4min 4s

    3DMARK01 SE
    2-2-2-5: 26,827
    2.5-3-3-7: 27,334

    Doom 3 1024x768 High Quality
    2-2-2-5: 95.2 FPS
    2.5-3-3-3-7: 96.0 FPS

    Doom 3 1280x1024 High Quality
    2-2-2-5: 75.0 FPS
    2.5-3-3-7: 77.8 FPS

    HL2 CS:Source Stress Test 1024x768 (Options maxed no aa/af)
    2-2-2-5: 144.82 FPS
    2.5-3-3-7: 145.71 FPS

    HL2 CS: Source Stress Test 1280x1024 (Options maxed no aa/af)
    2-2-2-5: 132.95 FPS
    2.5-3-3-7: 133.06 FPS



    I think this should help most of us out. In general, if you cant get high FSB from your ram to run 1:1 your cpu intensive applications are going to be slower (look at render and superpi, aswell as 3dmark01.) Higher FSB and bandwidth seems to help out quite a bit in windows and applications in general, but not so much in games. However, in games, even running a divider but @ 2-2-2-5 216fsb it pretty much on par with it going 42mhz faster @ 2.5-3-3-7.
    Comment from kristos @ 2005/05/24
    Quote:
    Originally posted by jmke
    or maybe consider buying MORE ram; as having more does not hurt, and you might be able to run a RAMDISK and increase performance tenfold
    I'm not sure what the problem was but I am sure there was an issue with the A64 and 4 sticks of ram.

    Like I said, I can't remember what. did it not work at all? did it only work for a lucky few? did it not work with 4*256 or 4*512 or both?
    Comment from Sidney @ 2005/05/24
    Venice and San Diego presumably can handle 4 sticks.
    Comment from jmke @ 2005/05/24
    the "problem" is that you need to drop the command rate to 2T instead of 1T, which will cause performance to drop a little;

    the newest A64 revisions do run 4 sticks at 1T
    Comment from jmke @ 2005/05/24
    Quote:
    Originally posted by Haut^Karl

    IMHO, I don't believe we "gamers" only play the most recent games and that's all. As you can see from the responses above, we re-encode DVDs & video and play many, many other games & applications.

    Here are Gamer's DVD Shrink performance numbers (the article has been updated with this also)



    that's ~4% performance increase going from CL2.5 3-3-7 @ 200Mhz to 280Mhz...
    Comment from Sidney @ 2005/05/24
    Comment from Bosw8er @ 2005/05/24
    Comment from kristos @ 2005/05/25
    Quote:
    Originally posted by jmke
    the "problem" is that you need to drop the command rate to 2T instead of 1T, which will cause performance to drop a little;

    the newest A64 revisions do run 4 sticks at 1T
    I have certainly not read every article but I remember from one that I did read that 4 sticks was still a no go.

    And I've seen tccd at 340'ish 2T get out performed by tccd at 270'ish 1T

    I don't know if any of this is a good representation of what we can generally expect with these settings/setups but I would like to see that proven wrong
    Comment from Sidney @ 2005/05/25
    Quote:
    Venice based CPUs have finally learned to work with four double-side memory modules supporting 400MHz.
    http://www.xbitlabs.com/articles/cpu...venice_14.html
    Comment from perdomot @ 2005/05/25
    Thanks for the update JMKE. Looks like getting tighter timings is almost as good as getting the HTT super high. Gives me something to think about.
    Comment from Sidney @ 2005/05/25
    I believe you could still get higher HTT whenever is allowed in your system; but you could use memory divider instead of getting the super expensive ones. I trust this is the very point JMKE is trying to express.

    Everybody has constraints and specially in dealing with money. Why spend $270 on memories in a system with <$150 CPU, like in my case, with slight improvement (a few %)in games?

    He indicated spending the money on newer HD, or graphic card would serve your system efficiency better.

    All bets are off if you have money to waste; or simply spending money pleases you more.
    Comment from Sidney @ 2005/05/25
    Comment from jmke @ 2005/05/25
    Quote:
    Originally posted by kristos
    I have certainly not read every article but I remember from one that I did read that 4 sticks was still a no go.
    I was wrong, with older CPU cores the memory speeds drops to PC2700/DDR333
    Comment from perdomot @ 2005/05/25
    JMKE,
    Well, I already spent some bread getting myself a Raptor and a 6600GT vid card so I think my rig will be plenty fast in those departments. I was wondering if you could test the effects of changing the different dividers available on the mobo without changing the multi/HTT. As an example, I am running at 255 X 9 and using the 133 divider to keep the ram below 200. Would getting memory that would allow me to use the 166 divider show any improvement in gaming or encoding? I don't recall any reviews that tested the effects of dividers on the performance of a rig. Thanks.
    Comment from jmke @ 2005/05/25
    Comment from Gamer @ 2005/05/25
    can't use dividers, don't know why....

    but you get the idea from the post above
    Comment from perdomot @ 2005/05/26
    Looks like the divider doesn't make much difference either. Guess the game is being able to get the highest megahertz possible.
    Comment from Sidney @ 2005/05/26
    No, it does not matter; and not much even with tight timing.
    Comment from GIBSON @ 2005/05/26
    Quote:
    Originally posted by jmke
    hey perdomot; CPU intensive tasks which rely solely (or mostly) on the CPU alone will show benefits; but this increase will not be that impressive either.

    I would advise to spend money on faster HDD, videocard, burner, etc, before buying speedier RAM; or maybe consider buying MORE ram; as having more does not hurt, and you might be able to run a RAMDISK and increase performance tenfold
    another interesting thing to do when you have 1G or more of ram is to load the windows core into your ram, it's a simple reg edit, from what i've heard it gives quite the boost
    Comment from jmke @ 2005/05/26
    have any links/docs on this? might be worth trying out and reporting the results
    Comment from GIBSON @ 2005/05/26
    well, i shouldn't be telling you this i think, but anyhow, a mate of mine made an optimised version of windows xp, i still have the reg file somewhere on my comp which has most of the reg edits, one of them in there is the one i talked about, i'll search it and i'll give you the edit, maybe something to be tested by [M] maybe it might help out on certain benchmarks too? (for those benchlovers out there )
    Comment from jmke @ 2005/05/26
    Keep the core system in RAM

    Windows XP takes portions of the operating system, applications, and data files that aren't currently needed in RAM and temporarily stores the data on the hard disk in the paging file. During a normal computing session, Windows regularly moves data back and forth from RAM and the paging file.

    If you have a considerable amount of RAM in your system--512 MB or more--you can improve system performance by preventing the OS from sending user-mode and kernel-mode drivers, as well as kernel-mode system code, to the paging file.

    You can do so by changing a setting in the registry. Here's how:


    Launch the Registry Editor (Regedit.exe).
    Go to HKEY_LOCAL_MACHINE\SYSTEM\CurrentControlSet\Contro l\
    Session Manager\Memory Management.
    Double-click DisablePagingExecutive DWORD value.
    Change the value in the Value Data text box from 0 to 1, and click OK.
    Close the Registry Editor.

    You may need to restart the system or log out of Windows XP for the change to take effect.

    Note: Editing the registry is risky, so be sure you have a verified backup before making any changes.
    Comment from GIBSON @ 2005/05/26
    yes, that should be it, anyhow, i've added a reg file that should do the trick
    Comment from jmke @ 2005/05/27
    that key was already set on this machine; can't say I ever noticed a "major" performance boost
    Comment from perdomot @ 2005/05/27
    That's just one of the tricks you can use to add a little extra "oomph" to XP. Things like disabling unnecessary services and cutting back on the eye candy also add a little bit more.
    Comment from Rutar @ 2005/05/27
    Does that work as well with Win2k?


    I thought I heard that MHz being important on the A64 also has to do that the higher frequency tightens the timings of the memcontroller (all timings). So if you haven't maxed out the memory with A64 tweaker there is an additional boost.
    Comment from jmke @ 2005/05/27
    don't think it will work for 2000
    Comment from GIBSON @ 2005/05/27
    oh well, all bits help i think the key isn't set by default though, unless you have a pirated version atm?
    Comment from jmke @ 2005/05/27
    I tweaked my Workstation's XP with this guide about a year ago: http://www.tweakhound.com/xp/xptweaks/supertweaks1.htm
    Comment from GIBSON @ 2005/05/27
    thnx for the link, i'll check it out
    Comment from perdomot @ 2005/05/29
    JMKE,
    Got a curve ball for you. I remember reading the article that showed that using a single stick of ram in SKT 754 rigs allowed for higher OCs than 2 sticks and since max mhz seems to be the key to performance, I was wondering if you'd run your bandwidth tests using 1 stick on a 939 rig and then compare the real world tests. I've got the sneaking sensation that you'd be able to get a higher HTT using one stick on the 939 rig and that would in turn allow for better performance since it would boost the total max mhz you could OC a PC to. If I'm wrong in my logic, please let me know and thanks.
    Comment from Gamer @ 2005/05/29
    you need about 150 mhz more to beat Dual Channel.
    Comment from Sidney @ 2005/05/29
    Its called "robbing Peter to pay Paul"; yes Paul is the bastard, but Peter remains wealthy after being robbed in this case.
    Comment from GIBSON @ 2005/05/29
    Quote:
    Originally posted by lazyman
    Its called "robbing Peter to pay Paul"; yes Paul is the bastard, but Peter remains wealthy after being robbed in this case.
    lol, nicely put lazyman
    Comment from perdomot @ 2005/05/29
    I'm going to test this out today and see what happens. If I can get back to 270 or more on 1 stick, I'll be ordering a 1 GB stick and running single channel. Hopefully, other people will run some benchies too. I plan on checking how vid encoding changes by testing with TMPEG and V-Dub to see if there is a difference between mpeg and divx. If anyone can do the usual Sandra stuff, we'll be able to get a good comparison.
    Comment from GIBSON @ 2005/05/29
    i wouldn't hope for it, i don't think you'll be able to make up for the lost bandwith with some extra mhz on your cpu,
    Comment from jmke @ 2005/05/29
    the 3400+ is a S754 setup in these tests: http://www.madshrimps.be/?action=get...&articID=23 0

    clock speed is all the same; to it comes down to pure core improvements, cache size and DC memory.
    Comment from wittekakker @ 2005/05/29
    Hey guys,
    I was wondering how big the difference would be in an enthousiasts benching program like super-pi.
    low fsb bad timings
    low fsb tight timings
    high fsb bad timings
    high fsb tight timings

    since I don't have A64 I can't do it myself.
    Comment from jmke @ 2005/05/29
    www.xtremesystems.org holds all the answer to that one
    Comment from perdomot @ 2005/05/29
    Update:
    I pulled out one of my sticks of ram and tested out the rig. I also tested with the two sticks in to serve as a baseline. I used two programs to encode an avi file into 2 different formats, mpeg and divx so as to see what differences there are. Here are the results:

    9 X 255 in Dual Channel Mode:
    Tmpeg: 46 seconds to encode a 1 minute file.
    VDub: 38 seconds to convert to divx.

    9 X 255 in Single Channel Mode:
    Tmpeg: 49 seconds.
    VDub: 40 seconds.

    9 X 270 in Single Channel Mode:
    Tmpeg: 46 seconds.
    VDub: 38 seconds.

    Observations thus far:
    1-Dual channel saves between 2-3 seconds per minute on encoding apps. This can save you between 4-6 minutes when encoding a 2 hour movie.

    2-Running the rig about 135mhz faster balances things back to DC level but I think there should be an increase in performance in other areas that would justify having the higher HTT in single channel mode. Hopefully, somebody how has Sandra and other standard benchies installed will be able to test this out.

    One possible fly in the ointment: The two 512MB sticks are from different manufacturers so I'm going to test the second stick to see if it is what limits me to 255 in DC mode. When I first set up the rig, I had 2 sticks from PQI and was able to run at 270 smoothly. I didn't think it would be a problem since the ram wasn't going to be run at full speed but I will test this next and report my results.
    Comment from perdomot @ 2005/05/29
    More info:
    I test the other stick alone in the rig and got the exact same results. Considering its quality samsung ram, I'm not surprised. This is making me think that the SKT 939 mobos benefit more from running one stick than two. Hopefully, somebody will run the tests for Sandra, etc. and post the results. The next step for me will be seeing how high I can take the HTT with the default multiplier of 9. Hoping for 2.5 Ghz
    Comment from perdomot @ 2005/05/29
    Even more info:
    Well, I decided to test moving the two sticks to channels 3 & 4 since I recalled some folks doing that and it seems to have worked for me. Am now at 9 X 270 in DC mode

    I ran the two encoding tests and saw a small gain. Saved one second on TMPEG and two seconds on VDub. Not much but I'll take any little bit I can get. Also found out that 270 is about the max HTT I can reach thats totally stable with the 9 multiplier. Hope this info helps folks out a bit.
    Comment from perdomot @ 2005/05/29
    Well, 270 started giving me problems even with one stick so I'm back to my theory that the quality of the ram makes a difference. The two sticks of PQI Turbo didn't give me any problems at 270 and I was using the 166 divider which made the memory run at 224. Even with the 133 divider, the cheap value ram seems to have issues so I think people thinking about the cheapy value ram might want to consider this.
    Comment from GIBSON @ 2005/05/29
    i think all of us here already knew that value ram wouldn't be able to hit 270HTT, even with those dividers
    Comment from perdomot @ 2005/05/30
    Why not? The dividers should have the ram running about 178 which is below its rated 200 so I don't see what the problem is. Can you explain?
    Comment from GIBSON @ 2005/05/30
    no, i wouldn't really have an idea why, maybe it isn't fully compatible with your mobo, but as i said, i wouldn't have a clue
    Comment from jmke @ 2005/05/30
    Quote:
    Originally posted by GIBSON
    i think all of us here already knew that value ram wouldn't be able to hit 270HTT, even with those dividers
    you're not making sense?

    if you are using dividers you let the ram run slower then your HTT, and the only advantage this can have, is that you can clock your CPU higher if you don't have an Athlon FX, which means you'll be stuck with a maximum multiplier.

    there is no other advantage of increasing your HTT and using dividers imho
    Comment from perdomot @ 2005/05/31
    JMKE,
    That's what I thought. The only two other possibilities that occur to me are:

    The PQI Turbo ram I could hit 270 with was single sided whereas the value ram is double sided. Since there is no Bank Interleave control in the bios of my VNF4, I think this might be affecting my OC potential. I might need to get some single sided 512 MB modules if this is the problem but I'm not sure if those are made.

    The only other thing that occurs to me is that maybe the PCB is causing problems but I am totally ignorant of how the PCB affects ram performance. Any suggestions?
    Comment from GIBSON @ 2005/05/31
    lol, indeed you are right jmke, duno what i was thinking when i wrote that it was kinda late so i prolly wasn't thinking very clear anyhow, what i prolly meant was what i wrote, without the part of the dividers, and permodot, i don't think they make singlesided modules of 512, they do with 256 though if i'm not mistaken
    Comment from perdomot @ 2005/06/01
    That's what I think Gibson which could be a problem then if the bank interleave is what is causing the issue.
    Comment from GIBSON @ 2005/06/01
    i guess switching to 2*256 isn't really a good option eh, 512 is not enough these days i think, btw, are you sure there aren't any bank interleave options? (sometimes motherboard makers hide certain advanced options to keep things tidy)
    Comment from perdomot @ 2005/06/01
    I've looked, believe me. Only other option is to get a different mobo but I'm not sure which one. This Chaintech VNF4 was pretty inexpensive, $85, so trying to find something that OCs well at a low price will be a little tricky. Any recommendations for an NF4 Ultra mobo?
    Comment from GIBSON @ 2005/06/01
    the DFI nf4 ultra should be good for oc'ing, it runs at +-130€ atm, it's a bit more, but i bet you get a lot more extra's too
    personally i'm planning to buy a DFI nf4 SLI-DR in july

     

    reply