Dual Memory controller slow on nforce2

[Stormseeker]

I have been able to get my system to run at 225MHz fsb on my Epox 8RDA3+ board, using 2.9v(really 2.82v) on OCZ 3500 ram dual low latency matched pairs, cas 2.0 timings 2, 2, 9, 2.0.

But each single stick of ddr 3200 when run at 200 mhzfsb should be capable of a theoretical 3,200mb/s.

In Sandra I get 2956Mb/s on a single stick in any dimm slot at 200mhz t-bred cpu at 11.0 ratio.

But Dual channel mode should be like raid 0 on hard drives, using 2 channels to transfer half the info to the 1st device and the other half to the 2nd device, whilst the 1st device recovers.

But In nforce 2 dual channel DDR mode I get a maximum bandwidth of 3101mb/s (11.5 x 200) at 200 mhz fsb.

This is a tiny improvement in bandwidth for using 2 devices, it does not even reach 100% of the theoretical maximum for a single device, let alone 2.

The reason I bring this subject up is because 2 of my friends have changed from nforce 2 / amd bartons, to Intel p4 2.4C (800 fsb version), using dual channel 3200 DDR.

Now they get Huge increases in bandwidth using 2 sticks.
ie from 3021mb/s in single mode on a i 875 chipset, to 4106mb/s.

Now If someone can change a nforce 2 bios And OR the nforce 2 memory controllers to make even 100% of the theoretical 3200 single channel bandwidth on a dual channel enabled system I would be happy.


BTW My system is Water cooled see my page:

http://mysite.freeserve.com/Stormseeker/index.html

Thanks for listening to my moans.

[Lord Vetinari]

Someone else will probably answer this better- but you are comparing apples with pairs.The AMD cpu can't use all that bandwidth (plus you are not allowing for memory efficiency it doesn't run at 100%).
AMD cpus shift data slower but do more with it than P4s. P4s have got to have high memory bandwidth to make up for this. So the whole structure of the MB chipset etc is tailored for them to do this. Likewise the various chipsets for AMD CPUs match the abilities of the CPU.

[SteveI]

The bottleneck is the FSB of the processor. P4 is quadpumped @200 mhz. Therefore, dual DDR @ 200 is more appropriate (2x 2x = 4x).. OTOH, AMD CPUs are double pumped, so only taking advantage of the DDR archetecture.

[Áedán]

Why? Well two channels of RAM really are faster, but the bottleneck becomes the FSB, not the RAM!

Remember that the P4 is using a quad pumped bus running at 100/133MHz, and the Athlon is using a dual pumped bus running at 100/133/166MHz.

Hence, adding another stick of RAM to a P4 system makes a difference, as the FSB is running at least twice as fast as the RAM!

On the Athlon system, the RAM is operating at the same speed as the FSB, so no huge benefits. Strange however, that the Athlon can still keep up with the P4.

Even stranger how Alpha machines equipped with PC100 RAM can outrun P4/Athlon machines with DDR400 RAM.

As always, faster memory does not necessarily translate into faster processing.

Also, the methods that benchmark programs use to measure the bandwidth doesn't often occur in the real world, so the real world benefit of running two channel is lower than the benchmarks would indicate.

[nfrenay]

Well, I think I can't add much to the answer. It's all very well explained.

I have imagined something that we probably won't be able to see, but imagine if some chipset + motherboard maker made a system where you could run VERY HIGH FSBs on athlons while using a dual-channel architecture and asynchronous memory bus.

Ex: 400MHz for the processor FSB (800 data effective)
while the memories are at 200-250MHz dual-channel.

That would be nice. I wonder how it would compare to actual systems and Pentium 4s.

Ah, just dreaming...

[SteveI]

nfrenay,
now that's a different question... Unfortunately, it wouldn't run fast. NF2 is not efficient in asych operations... A good way to test this would be to run asych with FSB at 200, and memory much slower in dual mode. I haven't tried myself, but I've been told that memory performance isn't very good.

[MiGGinZ]

Originally Posted by nfrenay

Well, I think I can't add much to the answer. It's all very well explained.

I have imagined something that we probably won't be able to see, but imagine if some chipset + motherboard maker made a system where you could run VERY HIGH FSBs on athlons while using a dual-channel architecture and asynchronous memory bus.

Ex: 400MHz for the processor FSB (800 data effective)
while the memories are at 200-250MHz dual-channel.

That would be nice. I wonder how it would compare to actual systems and Pentium 4s.

Ah, just dreaming...

If the quadband memory boards get off the floor the athlon XP will have this funcionality ALBEIT in a different socket format.

[nfrenay]

I know that SteveI, but what I meant to say is: "if" there was a new chipset+board that supported that (even if memory was fixed at half FSB for example, which is theoretically perfect for a dual-channel system on AMD), it would be nice.

When nForce2 was released that was my expectation. But as all AMD chipsets, asynchronous just kills speed.

Migginz: I didn't know about that...

[SteveI]

Originally Posted by nfrenay

...
When nForce2 was released that was my expectation. But as all AMD chipsets, asynchronous just kills speed.....

Ditto. The opportunities for inexpensive upgrades would have been tremendous.