NF7-S v2.0 & Dual DDR 333mhz Vs. 400mhz

Discussion in 'Abit' started by tHatDudeUK, Sep 8, 2003.

  1. tHatDudeUK

    tHatDudeUK Guest

    I have read a suggestion that running two 333mhz RAM chips in DUAL DDR mode
    can be as fast as 400mhz RAM as the chips are accessed at the same time
    hence maxing out at the 400mhz FSB. This would save me money should it be
    true as I already have a 512mb DDR333 chip and this would mean I would only
    have to buy another DDR333 chip rather than get rid of my existing 333 chip
    and buy 2 new DDR400 chips.

    Whichever RAM I use I'll be running at 200mhz FSB...

    Any thoughts or benchmarks would be very much appreciated.

    TIA

    tHatDudeUK.
     
    tHatDudeUK, Sep 8, 2003
    #1
    1. Advertisements

  2. tHatDudeUK

    tHatDudeUK Guest

    tHatDudeUK, Sep 8, 2003
    #2
    1. Advertisements

  3. Running dual channel memory with 200MHz FSB & memory at 167MHz (ie. 6/5) is
    slower than 200MHz FSB & memory at 200MHz (6/6) but not a whole lot slower
    in overall performance. I had also thought the same way as you since NVidia
    ought to be able to 'stripe' the memory like a RAID array to more than match
    the maximum CPU to chipset bandwidth using much slower memory with a low
    divider but it doesn't appear to work at all well.

    a) I was running one NF7-S & Tbred B 1700+ at 11.5x205.5 for 2363MHz with
    2x256MB Samsung PC2700 at 171MHz (6/5) & spec' CL 2.5-3-3-6

    b) I increased CPU speed to 12x200.5MHz for 2406MHz with the same memory
    also at 200MHz (6/6) with slower than spec timings CL 3-3-3-7

    (b) has a 1.8% faster CPU than (a)
    (b) has 17% faster memory than (a)
    (b) has higher memory latency than (a)

    Running SETI 24x7 I gained roughly 3% increased overall speed - 0.35 work
    units extra per day, not a lot but better than a kick in the teeth I guess.
    ;-)

    SETI is very sensitive to all three variables; Raw CPU speed, memory
    bandwidth/speed & latency.

    +++++

    I'm going to run some tests for you to see how much difference it makes in
    main memory bandwidth reported by Memtest86. [The Memtest86 v3.0 bandwidth
    figure seems to equate to 'real world performance' quite well & it's a very
    fast check compared to SiSoft Sandra.]

    NF7-S v2.0 NF7_d18 BIOS

    2x256MB Samung PC2700 - memory timings were left exactly the same for all
    tests - CL 3-3-3-7 The ratio at left is the FSB:Memory ratio as shown in
    BIOS.

    I tweaked the FSB a little to get the CPU speeds as close as possible,
    that's the 'why' of the 'Set' figure - you don't get the FSB you set in BIOS
    exactly. It means the memory speeds don't match exactly but.............

    A) CPU 16.5x131 for 2167MHz (Set 131MHz FSB, 2162 actual CPU.)
    Memory:
    A1) 4/3 (196MHz) 820MB/s
    A2) 4/4 (262MHz) 949MB/s
    A3) 4/5 (327MHz) 960MB/s
    A4) 4/6 (393MHz) 993MB/s

    B) CPU 13x166 for 2167MHz (Set 165MHz FSB, 2158 actual CPU.)
    Memory:
    B1) 5/3 (198MHz) 878MB/s
    B2) 5/4 (264MHz) 1040MB/s
    B3) 5/5 (330MHz) 1204MB/s
    B4) 5/6 (396MHz) 1204MB/s

    (Odd that those last two match _exactly_ but I ran multiple tests and the
    figures are repeatable.)

    C) CPU 11x197 for 2167MHz (Set 195MHz FSB, 2159 actual CPU.)
    Memory:
    C1) 6/3 (195MHz) 916MB/s
    C2) 6/4 (260MHz) 1095MB/s
    C3) 6/5 (325MHz) 1222MB/s
    C4) 6/6 (390MHz) 1454MB/s


    All 1:1 'synchronous' figures from above:
    A2) 4/4 (262MHz) 949MB/s
    B3) 5/5 (330MHz) 1204MB/s
    C4) 6/6 (390MHz) 1454MB/s

    That last set above show how much benefit in memory bandwidth there is from
    increasing FSB with memory in 1:1 synch'..! Overall performance for a
    memory hungry app' like SETI gets a huge boost too. 'Real world slowdown'
    compared to 11x197 was VERY obvious when loading & running Memtest86 whilst
    I had the multiplier up at 16.5x with the memory running at only 196MHz...


    All (close to) 333MHz memory:
    A3) 4/5 (327MHz) 960MB/s
    B3) 5/5 (330MHz) 1204MB/s
    C3) 6/5 (325MHz) 1222MB/s

    Summary:
    There's very little, if any, benefit from running memory faster than FSB- it
    figures in with the Athlon's single channel DDR CPU to Northbridge bus
    anyways. (Much the same for single channel DDR too.)

    There's a substantial loss in bandwidth (Circa 25%) running the Athlon's
    single channel FSB faster than the dual channel PC2700 memory, which is
    counter intuitive - if the dual channel memory controllers were even
    reasonably efficient then you ought to see hardly any loss at all methinks.

    Oh well, most chipsets seem to 'prefer' a synchronous 1:1 memory:FSB ratio &
    NForce2 appears to be no exception...

    +++++

    That lot took me an hour & 40 minutes - you owe me a SETI work unit. ;-)

    Ciao...

    [UK]_Nick...
     
    Nick M V Salmon, Sep 8, 2003
    #3
  4. tHatDudeUK

    tHatDudeUK Guest

    Thank you very much. I'm quite surprised at the results. I'm going to try
    and overclock my RAM stick anyway with a thermaltake DDR active cooler stuck
    on it. Then I only need another stick of RAM...

    Sorry but I only have a 486 DX4/100 base spare at the moment (will actually
    run at 120mhz with a faster hard disk :) ) I don't think you'd get too much
    Seti Action out of that :-S It will also cost me more to send it than I'd
    get for it on Ebay...
     
    tHatDudeUK, Sep 8, 2003
    #4


  5. Hi Nick,
    thanks for that detailed report. Sadly I'm not sure what the meaning of it
    is? What are you exploring here?. In a nutshell I always thought that to
    get the best performance out of an AMD platform you needs to run the fastest
    FSB/Memory bus (1:1) that you could get away with. Are you talking about
    the possible benefits of running Async with the memory up crazy high like
    250MHz-FSB/500DDR?

    Needed a conclusion to that post for simple folk like moi :p

    Wayne ][
     
    Wayne Youngman, Sep 9, 2003
    #5
  6. I've got a matched pair of 2x256MB Samsung PC2700 running at 200MHz with
    very relaxed timings. (that's the machine I ran the tests on.) Yours may do
    a really decent speed too......

    Heh - I was only kidding. ;-) You could do a work unit for me from there by
    using my email to login but a 486DX4/100 would take circa 10 days to do one
    SETI work unit - my farm of seven machines churn out circa 74 work unit
    results per day. :) Circa £50 a month in electricity bills with my
    single 'supercooled' Peltier rig costing about £20 a month out of that
    total..!

    I often do some form of test like that when I see, or think of, a question
    that interests me - you rarely if ever know all the parameters an online
    reviewer has used, so a personal test is often worthwhile methinks...

    Ciao...

    [UK]_Nick...
     
    Nick M V Salmon, Sep 9, 2003
    #6
  7. tHatDudeUK

    - HAL9000 Guest

    Nick,

    Thanks a bunch for taking the time to do all these tests. I think it
    was me that prompted tHatDudeUK post.

    What is significant to me about your test results are the A1, A2, A3,
    A4 results. My recollection is that a non dual channel board would
    have identical results in A2, A3 and A4 tests. That is, increasing
    the FSB beyond the memory rate results in (essentially) 0 performance
    gain - maybe even a performance loss because of non synchronicity.
    Apparently the dual channel configuration is giving you 993/949 = 4.6%
    performance gain with the 4/6 setting.

    That's sad, the dual channel conig isn't buying much. Perhaps it is
    just enough to tip the scales towards an nForce 2 for someone reading
    and comparing performance reviews though.

    Since I have a single channel board, I'll have to verify the A2, A3,
    A4 test as providing no gain on my system.

    I think I would leave the 4/6 setting and then try to maximize the FSB
    frequency. Then let the CPU multiplier be what ever gives you
    something stable.

    Thanks again,
    Forrest

    Motherboard Help By HAL web site:
    http://home.comcast.net/~hal-9000/


     
    - HAL9000, Sep 9, 2003
    #7
  8. No, the other way around - although I did run some of the tests that way
    around (Memory > FSB) just to show that there is little to zero benefit from
    it, especially with DC-DDR.

    The OP wants to re-use his single 512MB stick of PC2700 (333MHz DDR) memory
    with another one the same in his new NF7 - he theorised (& I agree with him)
    that dual channel memory at 333MHz *should* easily be able to handle the
    full bandwidth requirement of an Athlon at 200MHz FSB. ie. 2x166MHz DDR >
    1x200MHz DDR. I ran some tests to show that, although fine in theory, it
    doesn't work that way in practice - as you wrote above, pretty much any AMD
    system runs fastest with the memory bus & processor bus in 1:1 synch'.

    With the advent of the Opteron & Athlon 64 CPUs I suspect the NForce2 is the
    last Socket A chipset iteration from NVidia - pity 'cause I also suspect
    this aspect of a dual channel chipset would have sold a LOT of NForce
    motherboards if it worked at all efficiently - imagine being able to run the
    CPU FSB at 266MHz (533DDR) with 'old' PC2100 (266DDR) in dual channel with
    little to no reduction in memory bandwidth..!

    +++

    My NF7 system that has 2x256MB PC2700 memory is now running 200MHz 1:1 with
    the memory timings backed off - I gained 4% in overall system performance
    running SETI with the changes, although ~2% of that was an increase in raw
    CPU speed. That system is still 3% slower than another two systems in
    overall performance - they're all running at exactly the same speed (12x200)
    but the other two use Corsair XMS3200LL - that 3% stems from the difference
    in memory timings;

    1) 2x256MB XMS3200LL TwinX - 2-5-2-2 - WinXP Home
    2) 2x512MB XMS3200LL TwinX - 2-6-2-2 - WinServ03
    3) 2x256MB Samsung PC2700 - 3-7-3-3 - WinXP Home

    The second figure in the timings above makes little difference in perfomance
    but the other three figures all affect overall performance a little,
    especially for SETI - #3 is 3% slower for SETI than both #1 & #2.

    +++

    One weird thing though - #1 above is a Barton & #2 is a Tbred B - the Barton
    really ought to be quite a few percent faster for SETI with it's 512KB L2
    cache but it isn't - the difference may be due to the difference in
    operating systems. WinServ03 tested out to be exactly 3% faster than WinXP
    Home for SETI in some tests a friend ran on a computer he can multi-boot
    with a whole bunch of different O/Ss. The only other difference between my
    two systems is #1 is an NF7-S v2 & #2 is a vanilla NF7 v1.2 - that ought to
    make no difference, so it must be down to the operating systems I guess.

    Pity blasted WinServ03 costs so much - mine is from an 'evaluation CD' that
    will cease working in about six months. :-/

    ;-)

    [UK]_Nick...
     
    Nick M V Salmon, Sep 9, 2003
    #8
    1. Advertisements

Ask a Question

Want to reply to this thread or ask your own question?

You'll need to choose a username for the site, which only take a couple of moments (here). After that, you can post your question and our members will help you out.