The TCCD-pire Strikes Back? Athlon 64 S754 Memory Roundup

Memory by jmke @ 2004-11-30

We compared 5 different Samsung TCCD powered memory modules, discovering some surprising performance differences between them. To give these new TCCD chips a worthy competitor we also included results from Winbond BH-6 running with high VDIMM. How do they fair on an A64 S754 platform? Read on to find out

  • prev
  • next

Test Setup and Methodology

Setting up the Test System

Back in June I bought an MSI K8N Neo and AMD 3000+ CPU in order to do delve into the Athlon 64 world. With the advantage of being to able to lower the CPU’s multiplier I had good hopes for achieving higher FSB (or in A64 language: HTT) speeds to those obtained with my Asus P4C800 / P4 2.4 setup.

So I installed the system and removed the G.Skill PC4400 rated memory from its packaging and placed the 2 modules in memory slots 1 and 2. Booted up the system and went straight into the BIOS to drop the multiplier, relax memory timings to CL2.5 3-3-10, set VDDR to 2.8v and finally increased the HTT speed to 275Mhz. Then I rebooted, or rather, I tried to, as I was greeted with a blank screen every single time.

Dropping the HTT speed to 230Mhz allowed me to boot into Windows, hardly impressive I’d say. Some research into Athlon 64 overclocking showed several limiting factors for reaching high HTT speeds, beside the memory. At that time DFI released their now highly acclaimed enthusiast Socket 754 board, the DFI Lanparty UT NF3 250GB, so I ordered one and upon arrival of the board continued testing.

Much to my surprise did the system again fail to work stable at speeds near 275Mhz, using Memtest86+ it showed tons of errors, booting into Windows was impossible. So again I scanned through online forums and technical support pages, only to find the CPU as being the possible culprit for these instabilities and lack of OC potential. An Athlon 64 3200+ was bought and testing continued, when Memtest86+ again failed at speeds above 250Mhz I was getting desperate.

Before throwing in the towel I booted up the system with only 1 stick in place, lo and behold it started up without a hitch and worked perfectly. Other people discovered similar behavior with their socket 754 systems, getting good to very good result using only one memory module.

So is the A64 S754 CPU flawed and not able to run with more then 1 memory module? Surely not, at PC3200 speeds all memory modules performed excellent with very tight timings, it’s only when you push the HTT speeds to high levels that the on-die memory controller of the Athlon 64 will start acting up. Different steppings of the A64 also play a role in this, “C0” being known for awful performance, “CG” showing an improvement in this aspect. Unfortunately, even with the new A64 3200+ CG stepping, HTT speeds were impossible with those 2 G.Skill modules.

Yet there is still hope, as it seems to boil down to a compatibility problem with certain PCB (Printed Circuit Board) of the memory modules, which is the only thing that separates the modules from each other seeing as they are all using Samsung’s TCCD memory chips.

So this leaves us with PCB as the deciding factor.

Nathan Kirsch from Legit Reviews wrote an excellent article about the different PCB’s used by manufactures. They can be divided in three large groups:

  • those using JEDEC reference PCB
  • those using Brain Power’s B6U8** PCB
  • those using a modified JEDEC reference PCB design

    Nathan did his tests on an Intel setup and his results correspond with those I obtained in my Intel roundup. On Intel the modules with Brain Power’s PCB performed noticeably better and in today’s roundup I like to see if the same holds true for the A64 S754 platform.

    Furthermore, there is this post in our forums which shows a close up of 2 G.Skill modules using Brain Power PCB, the photo comes straight from G.Skill’s own Kevin Wu, who stated that in his own experience the B6U808 performed better then B6U815. These module numbers correspond to revisions of the Brian Power PCB. Do note that G.Skill rates their memory modules using an Intel setup, so whether or not the same “rules” apply for A64 you will find out pretty soon.

    In the roundup I have a mix of PCB’s and each one performed different on the A64 test setup.

    Test Setup and Methodology

    JMke's A64 Test Setup
    CPU AMD Athlon 64 3200+ S754 “CG” stepping
    Mainboard DFI Lanparty UT NF3 250Gb (Oskar Wu’s 10/15 Beta Bios)
    Cooling Stock AMD Heatsink and Fan
    Video nVidia Geforce 4 Ti4600

    The system is installed on an open test bed, for extra cooling an 80mm fan was placed pointed towards the memory banks.

    Madshrimps (c)

    The latest Beta bios for the DFI NF3 board adds Memtest86+ 1.27 as an option which was of course quite handy for testing memory stability. For finding the maximum stable overclock test #5 was looped for an extended period of time to check for errors;

    Madshrimps (c)

    In my testing I found that the maximum “zero-error” Memtest86+ speed did not always correspond with stability in Windows. And further stress testing was done by looping 3DMark2001, Prime95 and Q3A, each application is very sensitive to system overclocking and will fail quickly when errors are found.

    Each memory kit was tested using 1 and 2 modules, and next to finding the maximum FSB speed, performance was evaluated with 3DMark2001SE which serves greatly as an overall system benchmark, and also Quake 3 Arena, a game which depends on memory bandwidth for obtaining maximum FPS.

    Maximum FSB was searched while using 2.7v with these memory timings:

  • CL2 2-2-10
  • CL2.5 3-3-10
  • CL3 4-4-10

    The last setting was used for reference, as the performance will decrease drastically when using such relaxed timings. If the overclock potential did not increase by >10Mhz, then I did not include in the results.

    The DFI NF3 board offers a very wide range of memory timings which can be configured to improve overclocking, speed and stability. There is a very informative forum thread here explaining all the possibly settings, I did not tweak these settings to the extreme, as it would become impossible to finish this roundup in an acceptable timeframe, but I discovered 2 values which increased performance significantly with these TCCD modules:

    Madshrimps (c)

    DQS Skew Contol: Increase Skew
    DRAM Drive Strenght: Level 1

    These netted me an increase of ~10Mhz when overclocking at CL2.5 3-3-10. They did not impact OC potential at tight timings though.

    So with everything set for testing it’s time to show you the results ->
    • prev
    • next