Nforce2 Dual Channel Memory and Latency Benchmarking.

After hearing about some rumours on the internet about nforce 2 chipsets and memory latencies i decided to sit down and spend a few hours testing this. Now i have a Albatron KX18pro II, which is a nforce2 chipset with sata raid and a bunch of other neato things running.

I have 2 sticks of nanya ram running at 333mhz in dual channel mode. Now i have a bunch of pics that show my results in sisoft sandra and 3dmark. Now im not saying that you will get these results, and its not comparable to other machines, all im trying to do is test a few theories out.

One of the things im going to test out is the bandwidth equation. Now im not entirely sure about how Dual Channel mode works, as far as i can figure out, it doubles the theoritical bandwidth.

Lets say you have a 2600+ like i do. (its actually an unlocked 2400+)

It runs at a CPU-Chipset FSB of 333mhzDDR(166mhz).
I also have 512 meg of ram running at Ram-chipset FSB of 333mhzDDR(166mhz).
Now if my theory on dual channel is correct (in that it provides double the bandwidth) then this case would be equivalent of having 512meg of ram at 666mhz.This is reminiscent of the old DDR ram on pentium 3 situation where the RAM had double the bandwidth that the cpu could ask for. P3 fsb = 133mhz, DDR = 266mhz. There would be a waste of bandwidth. Now my test is this, IF this is true, then if i increase the FSB on my CPU, then it could theoritically use more of the available bandwidth of the ram. Now my CPU can only do a max of 400mhz DDR, (200mhz fsb), which isnt bad considering it WAS a 133fsb part.

So i propose a test. At 400mhz, will my available memory bandwidth be bigger then that available at 333mhz?

To test this i will keep my ram at 333mhzDDR(166fsb) while changing the CPU fsb between 333 and 400mhz.
If this test fails, i would have to ask, what is the advantage of DUAL channel mode when the CPU cant use it? I would assume that dual channel would help in that all memory bandwidth ISNT just used for the CPU, some is used in PCI transfers, some is used in IDE DMA transfers, some is used in AGP transfers. So i would at all times say that dual channel mode would always be a small amount faster then single channel mode. The question is, would the dual channel mode allow the cpu to suck more bandwidth? On a side note i will also test out the rumour that 2-2-2-11 is faster then 2-2-2-5.

Testing Procedure:
Using the incredibly detailed options in my motherboard bios, i spent 3 hours benchmarking.
Im using a radeon 9700pro overclocked. Im going to use 3dmark01, so its not so video card limited and more limited to cpu and ram. I am not going run it at low res, cos no one games at 640x480. So 1024x768 is a representative res. I changed bios settings, and when in windows ran the sisoft memory benchmarks and then 3dmark, i think rebooted, changed the settings and did that cycle. I decided not to run 3dmark 3 times, because it would have taken ages, as it is, it took 3 hours. (with a few alcoholic beverages to help pass the time)

I have a few options i have tested out.

RAM latencies set at 2-2-2-11 and 2-2-2-5
CPU interface set at "Optimal" and "Aggressive"
CPU fsb at 333 and 400.
RAM fsb locked at 333 with the variable divider set to 2.00 for 333fsb, and 1.66 for 400fsb to keep the ram always at 333mhz.

I will test with Sisoft sandra and 3dmark 2001 for comparative results.

Test Results at 2-2-2-5:

Test Results at 2-2-2-11:

So what does all that mean? Ok. Fastest 3dmark there is aggressive settings, 333 cpu, with ram at 2-2-2-5. Next fastest is same settings but with ram set at 2-2-2-11.

Conclusions:

At the 400mhz CPU fsb, integer bandwidth approximately increases across the board (ignoring the 2584/2583 which i will exclude due to measurement errors.)
The increase is so slight, but definately an increase, about 5-15 meg a second.
However, the float bandwidth decreases across the board, and by a larger margin then the increase of the integer side. Taking into account that games use a lot of floating point, i would take a guess at leaving the FSB at 333 for gaming. The average float bandwidth reduction is about 30-50 meg.

Another note is that the aggresive settings definately increase all the results, so i recommend leaving that option to aggressive, though this is a board specific setting.

Final note is the difference between 2-2-2-5 and 2-2-2-11. In same cases the 5 beats the 11, and in others its the other way around. The final word goes to the 3dmark, with the fastest score at 2-2-2-5. All in all its all pretty close, however the 16 grand mark is only beaten on aggresive settings, with the next closest being optimal at 333mhz with 2-2-2-5. At no time does a FSB of 400mhz break the 16000 barrier, with average scores being lower then when set at 333.

On an interesting side note, as well as the memory bandwith tests i did the Cache and Memory benchmark. The results are interesting, especially on the 333/400 side of things. Here is the 333 test on aggresive settings with 2-2-2-5.

And the same test on with 400mhz fsb....

Not so much difference until you get to the 512k block size which excludes the CPU l2 cache. Notice the 256meg block transfer. At 400mhz its 1439 meg a second, while at 333, its only 1374 meg a second.
Using the difference of 333/400 (being 120 percent increase) the 400mhz figure is only about 104 percent faster. However it is definately an increase and worthy of note.

Final Words:

Now the theoritical bandwidth of a 333mhz DDR data channel running at 32bit is 2656 meg a second. At 400mhz that figure increases to 3200. My theory of the Dual Channel Mode was that it provided a theoritical bandwidth of 5312 meg a second, due to the fact that its still running at 333mhz, but double the bit width (64bit instead of 32bit). Now having that available at the chipset, i would assume that running my CPU at a FSB of 400 should at least give it a decent boost, up along the lines to the 3000 meg a second mark. This has been proven NOT to be true. In the sisoft sandra benchmarks at 400fsb, it noted that the bandwidth efficiency is approximately 82 percent. At 333fsb, it said 97 percent efficient. The differences in memory bandwidth are miniscule.

So... what exactly is dual channel mode???

I dont have the patience to test this machine with only one stick of ram to test dual modes advantages over single channel mode. I might do it later this week, but i hope to have at least made you think about what exactly dual channel mode sets out to achieve, and if it works as intended in theory at least.

At this point of time, i have to scrap my theory of dual channel mode providing double the bandwidth, however i cant escape the knowledge that increasing the CPU fsb while keeping the ram speed the same DID in fact increase my bandwidth in at least the cache/memory benchmark of sisoft sandra. Maybe this points to some slow downs in the nforce 2 chipset or some other effect that i cant figure out. Or maybe, and this is what i think now, the dual channel mode provides lower latencies, which allows the cpu to address data quicker, and this is what brought the figures up.

Well i hoped to have helped... all i can really say from this whole experiment is...

Get dual channel mode, it can only help. With amd CPUs, im quite sure a lot of them can do 400mhz FSB(DDR) and will definately benifit from it, however if you are running slower ram, try and keep the ram and cpu fsb locked together. If you cant unlock your CPU yourself, a local geek can do it for you, or you can send your CPU away to be unlocked. The memory bandwidth available increases dramatically running 200mhz instead of 166 or 133 mhz fsb.

Hope it was a good read.

Rainwulf. 4Jan04