Quad chanel Memory !!!

http://www.warp2search.net/article.php?sid=10859&mode=thread&order=0

Nothing against DDR but I want that Chipset for my next PC
probly gona be a while a long while !!!

I'd like to see actual products being fast and stable and with good availability though. SiS is doing well lately, but they have had a poor track record in the past.

It's good to see this performance initiative though. I hope it materializes, together with speedy FSB's.

Hmm, quad pump RDRAM sounds good, but for the moment, the fastest P4 doesn't require 9.6 G/s bandwith, but it does require 6.4 G/s and that's precisely what the quadpump DDR3200 will provide (4 x 200 Mhz = FSB 800) with the i875 chipset and the D-1 core revision of their P4.
This chipset and P4 3Ghz D-1 are to be available this April, according to Intel roadmap. That probably means that we won't be able to order any of these before June . That's still a 4 months minimum wait, but it's still much better than if you wait for the new SiS chipset. Anyway, before that, the new ATI R350 GPU is coming so you have an excuse to spend money for a new video card, and also a pair of DDR3200.

FIZZ3 said:

It's good to see this performance initiative though. I hope it materializes, together with speedy FSB's.

Click to expand...

Currently, in P4 running quad speed FSB, as I understand, the max memory bandwdith is

maxMemoryBW = FSB x 4 x 8 = 32 FSB MB/s

e.g. FSB = 133 MHz, maxMemoryBW = 4256 MB/s
e.g. FSB = 150 MHz, maxMemoryBW = 4800 MB/s
e.g. FSB = 200 MHz, maxMemoryBW = 6400 MB/s

In practice, can someone w/ a P4 board tell me how much bandwidth in MB/s one can actually get by running SiSoft Sandra 2003 memory bandwidth benchmark at a certain FSB speed, if better, at different FSB, e.g. 133, 150, ... MHz.

Since I only have an AMD dual channel, I cannot measure it for P4 and would like to know how a P4 memory controller performs, and its efficiency compared to nforce2 dual channel.

Well you can't even compare Dual Channel DDR on P4's and AMD's, because the AMD's don't have anywhere near the kind of bus bandwidth it takes to utilize a dual channel setp.

But anyway, I'm running a dual channel setup on my system (in sig). At 166fsb with really slow timings (my memory is kinda cheapo ) I get a bit over 4GB/sec of memory bandwidth in Sandra. I think it's about 4050MB/sec.

Can't wait for my 1 gig of Corsair XMS to get here!

Stumpjumper5200 said:

Well you can't even compare Dual Channel DDR on P4's and AMD's, because the AMD's don't have anywhere near the kind of bus bandwidth it takes to utilize a dual channel setp.

But anyway, I'm running a dual channel setup on my system (in sig). At 166fsb with really slow timings (my memory is kinda cheapo ) I get a bit over 4GB/sec of memory bandwidth in Sandra. I think it's about 4050MB/sec.

Can't wait for my 1 gig of Corsair XMS to get here!

Click to expand...

What ratio of memory to fsb is usually used for P4? What fsb and memory speed are you running at? What is the max fsb for P4?

I know that in absolute term, Intel P4 has a much higher memory BW than AMD nforce2 dual channel.

Assume P4 max at 166 fsb, max BW = 166x4x8 = 5300 MB/s. While AMD nforce2 fsb max at 200 MHz, and max BW = 200*2*8 = 3200 MB/s.

I am try to compare the memory controller efficiency between Intel P4 and nforce2.

Since you run 166 fsb, the maxBW = 166x4x8 = 5312 MB/s and you are getting about 4000 MB/s. So do you agree that P4 memory efficiency ~ 4000/5312 = 75%.

For nforce2 whose absolute BW is much lower than P4 chip set. But I estimate its dual channel memory efficiency ranges from 50%-80% too. 50% is when running dual channel SYNC, since half of BW is wasted. 80% is when fsb is running much higher than memory speed (using slow memory).

E.g. fsb = 200 MHz, memory speed = 150 MHz, the dual channel effective BW ~ 2900 MB/s. So nforce2 memory efficiency = 2900/(2400*2) = 60%

fsb = 200 MHz, memory = 100 MHz, effective BW ~ 2400 MB/s. Efficiency = 2400/(1600*2) = 75%. It is more efficient to run at slow memory, but of course the actual BW is much lower.

I agree, I've calculated my efficiency to be about 75% as well before. This is also with slow timings, when I get my new memory, it should kick up a tiny bit.

Also, CPU:Mem ratios are not adjustable on the Granite Bay, because it's pointless to change it from its 1:1 ratio. So at 133fsb, each stick is doing 133MHz (DDR266), or basically Dual Channel DDR266. That works out to the theoretical 533MHz (2x266) of the bus at 133fsb.

Perfect

Stumpjumper5200 said:

I agree, I've calculated my efficiency to be about 75% as well before. This is also with slow timings, when I get my new memory, it should kick up a tiny bit.

Also, CPU:Mem ratios are not adjustable on the Granite Bay, because it's pointless to change it from its 1:1 ratio. So at 133fsb, each stick is doing 133MHz (DDR266), or basically Dual Channel DDR266. That works out to the theoretical 533MHz (2x266) of the bus at 133fsb.

Perfect

Click to expand...

So Granite Bay is basically a dual channel with fsb (nominal at 133) running at quad speed of memory speed (nominal 133, DDR 266).

Can the fsb be overclocked to say 166 MHz? If not, then there is no point of using memory faster than DDR 266. Is this true?

If it can be overclocked to 166 MHz, then DDR 333 would make sense to be used.

Can Granite Bay get close to 100% efficiency on memory BW? I.e. if memory run at 133 MHz (DDR 266), dual memory BW = 266 x 2 x 8 = 4256 MB/s. If measured by SiSoft, would you see memory bandwidth close to 4256 MB/s.

PS: when nforce2 is overclocked hard to 220 MHz, its fsb memory bandwidth = 220 x 2 x 8 = 3520 MB/s, which is within 20% from the Granite Bay nominal bandwidth.

Well yeah, it can be overclocked. In my case, I'm at around 166fsb. They also have locked AGP/PCI, which is awesome

And no, you can't get 100% efficiency. The memory is running fast enough to do so, but that's only in theory. You'll never be totally efficient, becuase the bus has other things to transfer than just the memory. You'll always have the controller overhead + other pieces of data the bus is transferring.

Stumpjumper5200 said:

Well yeah, it can be overclocked. In my case, I'm at around 166fsb. They also have locked AGP/PCI, which is awesome

And no, you can't get 100% efficiency. The memory is running fast enough to do so, but that's only in theory. You'll never be totally efficient, becuase the bus has other things to transfer than just the memory. You'll always have the controller overhead + other pieces of data the bus is transferring.

Click to expand...

What do you think how high the fsb of Granite Bay can be overclocked to? Is 166 more or less max?

If fsb 166 is max, and you mentioned dual channel efficiency is around 75%, the effective bandwidth of Granite Bay is around
166 * 4 * 8 *0.75 = 5312 MB/s * 0.75 ~ 4000 MB/s
All you need for Granite Bay now is DDR 333 memory (unless paying for the future).


For nforce2 @ 220 fsb, max bandwidth = 220 * 2 * 8 = 3520 Mb/s
Actual is around 3300 MB/s

So nforce2 is only about 20% behind Granite Bay in memory BW, but much cheaper for AMD CPU.

Who knows......if I could somehow get an engineering sample of a P4, maybe I could find out how high the fsb will go

The board has settings for 355fsb.