One of the more vexing discoveries as memory performance testing was extended to the Athlon 64 platform was that memory often did not perform the same on both platforms. You simply could not assume that a memory that did DDR550 on the Intel 478 chipset would perform the same on the Athlon 64. We have also discovered that certain memory actually performs much poorer on Athlon 64 than on Intel 478, while other memory consistently performs much better on Athlon 64. With these variations, it was time to establish an Athlon 64 memory test platform and run new baseline tests on some of the best recent memories.

You can see the details of our new Athlon 64 testbed in the recent review, OCZ 3700 Gold Rev. 3: DDR500 Value for Athlon 64 & Intel 478. The decision was made to look to the future on Athlon 64 memory benchmarks with a Socket 939 Dual-Channel test bed. With the recent introductions of the 90nm Socket 939 3000+, 3200+, and 3500+ processors, the starting price for a 939 CPU is now well below $200. This will likely encourage further growth of the 939 as the top-performance platform for Athlon 64. We will soon be bringing you performance and overclocking tests of the new 90nm AMD chips, and our decision to concentrate on Socket 939 for our Athlon 64 memory test bed was influenced by AMD's targeting of the 939 for new product developments.

VIA just launched their first reference boards using PCI Express on the Athlon 64. Later this month, we also expect new Athlon 64 chipsets from others that will add PCI Express capabilities to the Socket 939 platform. While these new chipsets could migrate later to Socket 754 single-channel, the new chipsets will launch with Socket 939. This will further push the 754 to the value side of the Athlon 64 line.

As you saw in our DFI LANParty UT nF3-250Gb: Overclocker's Dream review, the 754 is capable of incredible performance. It is even capable of outperforming the newer 939, since base performance is only about 5% higher for the 939. However, this usually requires the use of one DIMM. Overclocking performance with 2 DIMMs on 754 is normally poorer than 2 DIMMs on 939. While there are many reasons to buy 754 for value and performance, future development will revolve around the 939 socket and dual-channel memory.

To understand better how memory behaves on the Athlon 64, we tested a cross-section of some of the best memory currently available in the lab. This included new Samsung TCCD memory form PQI and G. Skill, familiar Samsung TCCD from Geil and OCZ, top performing Micron-based Crucial Ballistix, and the latest Hynix DT-D5 memory from OCZ. We had originally planned to include the unique OCZ 3700EB also, which had performed well in other Athlon 64 tests. However, OCZ told us EB memory was no longer in production, and we could not find EB in stock at any vendor. We, therefore, eliminated EB from our final testing, since it is no longer available for purchase.

Crucial Ballistix PC3200
Comments Locked


View All Comments

  • Wesley Fink - Friday, October 1, 2004 - link

    #22, #25 and others -

    The 465W power supply that was not up to the job had the following specifications:

    +3.3V - 38A
    +5V - 44A
    +12V - 20A
    -5V - 2A
    -12V - 1A
    +5Vsb - 2.2A

    I would never have had any reason to suspect issues with this expensive major brand PS based on those specs, but in fact it turned out to be the limiting factor in overclocking the memory.

    This issue requires more investigation as there are many possible reasons for my experience, but I felt an obligation to let readers know what we found in our testing. However, it is not fair to name brands without much more evidence.
  • Blappo - Friday, October 1, 2004 - link

    Hopefully the value memory roundup will also contain some generic memory as well. Knowing which premium memory is the faster is good, but I want to know how much of a difference memory makes on system performance. I don't want to spend 50% more for only 5% more performance. That money could be better spent somewhere else.

    I'm glad that AnandTech did this article since everywhere else they only benchmark memory on Intel systems.
  • Blappo - Friday, October 1, 2004 - link

    PrinceXizor - I agree completely. However, with higher wattage PSU most of the extra power comes on the 12V rail. I think that is why nVidia is recommending "monster" PSU for their high-end graphics cards because it is easier to tell people to buy a 480W PSU rather than a PSU that can deliver 22A (or whatever) on the 12V rail.
  • Spearhawk - Friday, October 1, 2004 - link

    15: 20A? You're kiding right? My oven are at 20A.
  • decptt - Friday, October 1, 2004 - link

    My Ballistix performance
    "Dec" copyright since Duron600@1000
    Athlon64 Mobile 3200+ Rev.SH7-CG
    AMA3200BEX5AR-ClawHammer L2:1MB
    10x250 vid:1.5v+113%(~1.70v)
    Idle@36C Load(Prime95)@43C
    [10x255 worked @~1.97v Loaded(Prime95)@55C]

    Crucial Ballistix PC3200 512MBx2 vcore:2.70v
    Ratio = 1:1 ; TCl:2.5-Trcd:3-Tras:5-Trp:3 1T
    ATI AIW 9600XT 128MB BUS:75 vcore:1.6v
    DFI Lanparty UT NF3-250 Rev.A00 Bios.9/14(Beta)
    LDT/FSB@ 4X
    Thermalright XP-90 +Panaflo-H1B-92(FBA09A12H)
    Seagate ST3120026A
    DVD model DD0401

    Scroll(Tested@2.50GHz on Sep 30, 2004)
    CPU::ALU 11513 FPU 3954 iSSE2 5156
    Media:: Int 23895 Float 2562
    Ram:: Int 3780 MB/s Float 3779 MB/s
    3DMark03V3.4.0--1751 :(
    Super PI 1M 35s
    Super PI 2M 1M23s
  • quanta - Friday, October 1, 2004 - link

    Speaking of robust 12V rails, Enermax seems to be the only one make power supplies that can provide more +12V juice than most enthusists ever need, even for models without splitting +12V lines.
  • AlphaFox - Friday, October 1, 2004 - link

    where did the power supply talk come from??
  • PrinceXizor - Friday, October 1, 2004 - link

    Just a further comment on the PSU "recommendation". I really think that the major tech sites are doing a disservice to the community when they keep recommending higher and higher total wattage PSU's. That is NOT the issue! As has already been pointed out, the key component is to analyze what voltage rail is is not providing enough juice. Just as clock speed is a poor indicator of processor performance, total wattage is a poor indicator of PSU performance. Considering that a major computer rig will rarely if EVER draw more than 250W of actual power, the key metrics for a PSU are the actual amperages on the various rails, particularly the robustness of the 12V rails. (Some newer PSU's are providing dual 12V rails for just this reason). So, does an enthusiast overclocking their rig need a 500 Watt "monster" or do they need a robust 12V line and tight voltage variance on those lines (as you drive stuff out of spec, the transients in those lines become more critical).

    I guess the point of my long-winded post is this...tossing off a "recommendation" like was done in this article (well-intentioned I'm sure) without addressing the actual issues involved seems to me to be habit that should be avoided.

  • rjm55 - Friday, October 1, 2004 - link

    As you pointed out, you can't do the kind of tests you did here with an Intel rig with a locked multiplier. It was good to see all the different memory speeds at the same CPU speed, which proved once and for all that higher memory speeds DO improve performance - even with slower timings at higher speed. Makes my mouth water thinking about how good DDR550 at 2.5-2-2 would be.

    I realize the performance differences weren't huge with just the memory overclocked, but most people will overclock the CPU AND the RAM and that will make a huge combined difference in performance. Tweaking is about squeezing the most from your gear, and you CAN get more out of memory at higher speeds.
  • eetnoyer - Friday, October 1, 2004 - link


Log in

Don't have an account? Sign up now