The myth of dual channel..

DogSoldierDogSoldier The heart of radical Amish country..
edited February 2004 in Hardware
Since I'm getting 2 sytems based on it, I wanted to know more about the Via PT800 chipset. There isn't much on the net but after reading these articles, I can understand why:

The Dual-Channel Memory Performance Myth--Mostly Sizzle, Not Much Steak.
and The Dual Channel Memory Performance Myth, Part II: VIA to the "Rescue".

The PT800 is actually a very fast single channel DDR chipset, with performance somewhere between the Intel 865 and 875. I'm looking forward to getting my first new system on thursday. Theoretically, it should be able to outperform my current system.

[H}ardOCP has a good article, http://www.hardocp.com/article.html?art=NDk1

Comments

  • croc_croc_ New
    edited February 2004
    Thats pretty interesting. I wonder how running single channel on the P4C800 and your P4P800 will affect the benchmarks compared to dual channel.
  • edited February 2004
    Sounds good...and I can remember back when I had more points than you...fold on!
  • JustinJustin Atlanta
    edited February 2004
    It would seem to me that dual channel would almost always outperform single channel just as a RAID-0 array would outperform a standard IDE drive. Am I wrong in this?
  • mmonninmmonnin Centreville, VA
    edited February 2004
    Well lets see, for a P4C the memory bandwidth should be cut in Half!
  • croc_croc_ New
    edited February 2004
    mmonnin wrote:
    Well lets see, for a P4C the memory bandwidth should be cut in Half!

    I don't mean bandwidth scores, I know that :P .... I mean performance benchmarks :)
  • GargGarg Purveyor of Lincoln Nightmares Icrontian
    edited February 2004
    DogSoldier wrote:
    Since I'm getting 2 sytems based on it, I wanted to know more about the Via PT800 chipset. There isn't much on the net but after reading these articles, I can understand why:

    The Dual-Channel Memory Performance Myth--Mostly Sizzle, Not Much Steak.
    and The Dual Channel Memory Performance Myth, Part II: VIA to the "Rescue".

    Wow, those Sudhain articles were good. Been awhlie since I've sat through a whole article on a tech site, much less two :D
  • DogSoldierDogSoldier The heart of radical Amish country..
    edited February 2004
    hahaha, so Croc.. did you do any benchmarks? And if you did.. how do you...uhm, turn off dual channel?! LOL
  • edited February 2004
    Just move both sticks of ram over to the same channel instead of running one stick in slot-x on channel 1 and the corresponding slot on channel 2.
    When they are both on one channel it will force the motherboard into "virtual single channel mode".
  • EyesOnlyEyesOnly Sweden New
    edited February 2004
    If i'm not completely mistaking p4s need dualchannel or they'll starve. athlons don't support it but the motherboards do. Makes you wonder if socket 939 will be faster than 754. The only known differens between the 2 are that 939 is supposed to support dualchannel. It's wait and see and less then 2 months left.

    Another edit.
    Looks like dualchannel isn't a must with the right chipset. Read the hard ocp article to see what i mean.
  • mmonninmmonnin Centreville, VA
    edited February 2004
    Dual channel is only needed when you are sending more data than one channel can handle. So some benchmarks wont benefit because it doesnt tax the bandwidth.
  • EyesOnlyEyesOnly Sweden New
    edited February 2004
    Yea but those were real world benchies like ut2k3. And that's what's important. How many people really bechmark their systems. Most do by plaing games and notecing if it runs slow or not. Not scientific but it works.
  • Straight_ManStraight_Man Geeky, in my own way Naples, FL Icrontian
    edited February 2004
    EyesOnly wrote:
    If i'm not completely mistaking p4s need dualchannel or they'll starve. athlons don't support it but the motherboards do. Makes you wonder if socket 939 will be faster than 754. The only known differens between the 2 are that 939 is supposed to support dualchannel. It's wait and see and less then 2 months left.

    Another edit.
    Looks like dualchannel isn't a must with the right chipset. Read the hard ocp article to see what i mean.

    You are correct in both cases, but Opteron is multipipe. If memory mfrs cannot get faster ram to handle the volume of a multipiped CPU flow, then we need multiple channels, independent of each other as to storage. Can each pipe in an Opteron be seen as a CPU??? Yes, lots of software sees each pipe as exactly that. Can Prescott run in a mode where each pipe is a CPU to 32 bit software??? Yes. How WELL, is yet to be seen.

    Given faster processors, RAM that cannot be made lots faster needs to be one independent channel per CPU(or in the case of 32 bit code, per virtual CPU, per pipe). Now, lets look at benchmarks-- how many benchmarks are coded to work BOTH pipes of a Prescott or the muliple pipes of an Opteron simultaneously?? VERY few, the scientific basis for benching is being worked on with scientific bench code now.

    BTW, Opteron 844's are available in US, price $2,300 dollars to nearest hundred (price I saw recently for one of those shipped was $2,274.00 plus or minus a buck for FedEx saver shipping). That is an 8-pipe on die CPU. Code made for it could thread 8 threads of 64 bit code simultaneously.

    The 32 bit code assumed a single CPU, with threads in different processing states in CPU. Now the 8-piped CPUs are out and available. If RAM does not catch up, expect quad-channel or greater.

    The current 32 bit code will be processed through a single pipe of fast CPUs. Code will take a couple years to perk fully to the place it needs to be to exploit an 8-pipe CPU (each pipe is like a full CPU ALU and FPU, excluding caching except fro relatively minimal cache neds to drive th pipe-- L3 cache can be assinged in blocks per pipe, L2 in engineering sense is per pipe and should NOT be shared, though if big enough it could be blocked and part of each L2 bank reserved for each pipe).

    Current benchmarks cannot take real advantage of what they do not know. threeading has to be desinged into how software works, so does parallel porcessing, which is what these CPUs can do if the code is set up from bottom up for that. The compilers for 4 pipe code are being worked on. In essence we are moving into an age where abox can have and will have one die with almost 8 complete CPUs in it-- due to caching and RAM limits we will not be there for a while, and dual channel is a transitional (memory) bus way to let curent RAM not get saturuated by providing multiuple independent banks, one per pipe, eventually. The IC7-Max3 I have now as motherboard (not installed yet, here though) can run standard RAM in both banks at JEDEC Dual-channel timing or very close, because there is one channel per separated bank. That is what is going onto the IC7-Max3 when I stick it in, and we will see how it folds that way.

    ther new CPUs have nto been out long enough for coders to find out what WILL reliably work and take best advantage of the new gen CPUs, so the old ones will be off compared to what the CPUs can do with code intended to use the unique features in the new hardware. RAM will DRAG on CPU, reducing performance, as CPU has to wait while RAM catches up-- if it doesn't, it saturates RAM by feeding data in too fast. One RAM channel per pipe is best way until RAM mfrs catch up, and we do not have that support hardware at reasonable prices yet for 4-pipe Opterons, much less 8-pipes in one CPU.

    Botom line, is that I hang 2-3 years behind hardware becasue software was developed on average 2-3 years behind the hardware-- and I am talking about optimized software. It takes that long as the compilers and machine code have to be worked out first to take advantage of what the CPUS now available can offer, then the programs are written using the dev software and compilers. The dev base has to be there first, then the applications can be written.

    Clusters are used now as each blade is a computer, and can handle older code. The more blades, the more parts of code can be worked on at once with one blade handling each part. Industrial Light & Magic uses clusters for VR-style game development as many modules of code can be run at once. Standard XP can be run on one or two CPUs, and if the CPUs can run each pipe like a CPU as far as the code is concerned, then you can have the CPU run two threads at once. But they only partly do this now, becasue of caching and how it is mapped to pipes of CPU.

    John D.
  • DogSoldierDogSoldier The heart of radical Amish country..
    edited February 2004
    Fascinating stuff, I wonder how fast that Opteron 844 folds. ;)

    I still think the Via PT800 is a quality single channel solution for 800fsb HT enabled procs, but they OC like ass. So I deepsixed the MSI board for a P4P800 ($40 more) and of course, this means I'll need some quality RAM, bleh...

    "I am in blood

    Stepped in so far, that, should I wade no more,

    Returning were as tedious as go o'er."
  • croc_croc_ New
    edited February 2004
    EyesOnly wrote:
    Yea but those were real world benchies like ut2k3. And that's what's important. How many people really bechmark their systems. Most do by plaing games and notecing if it runs slow or not. Not scientific but it works.

    Yeah, I plan to test it out with games AND synthetic tests. I can get higher FSB/lower timings on my ram going single channel.

    Dual Channel will probably be faster, but it will be interesting to see how much faster.
Sign In or Register to comment.