Overclock.net › Forums › Components › Memory › Where is the bottleneck in RAM access?
New Posts  All Forums:Forum Nav:

Where is the bottleneck in RAM access?

post #1 of 2
Thread Starter 
I am still an amateur when it comes to this stuff but I feel I should post my findings even though its not fully tested, which was the reason for the delay as I could have released this info years ago. And if true, its quite important for you to know. It seems a little funny this forum has separate categories for “intel memory” and “amd memory”, I shouldn’t double post, but I am sure word will get around.

When I was designing my computer years ago, I noticed something funny about the ram. While the speeds increased, the latencies increased as well. They were being sped up and then slowed down at the same time. This made me think the chips might have been exactly the same. I like to use spreadsheets so I mapped it. And yes, there was a common denominator, and I was correct. This was great because I could choose the slowest ram as it was cheaper.

CAS+RCD+RP+RAS is, I am guessing a sequence of operations requiring a certain time in nanoseconds to perform their operation. Increasing the megahertz frequency, requires an increase of clock cycles (latency) to readjust to the ‘certain time’. I am not sure if the other numbers are sequential with the first (RC RFC RRD WR WTR RTP) or if they are responsible for some of the nonlinear product ratings when getting to lower latencies/speeds (latencies like 3-4-3).

In my spreadsheet, I rated the memory based on the difference between latencies and the megahertz. The DDR mhz (eg: 800, as opposed to 400 for the actual ram freq) was divided by 133.3333….. (or 133+1/3) and I got an integer. This was divided by the sum of the three latency digits and this gave me a % value which I graded the ram on. Usually I got a rating of 44.444…%, slowest was 33.333…%. I collected overclockers results and different products. Best overclocker score I found was ‘the stilt’ getting 90% with D9’s at 4-2-2, 960mhz. These calculations still work with DDR3, so we can compare DDR2 with DDR3. Since the latency figures are integers and do not match well with the frequency so you will see some of the %’s are not quite the same (eg, 44.444 and 41.667, 55.55 and 50).
Product examples are:

(55%)2GB kit (1GBx2) BL2KIT12864AA663 DDR2 PC2-5300 3-3-3-12 UNBUFFERED NON-ECC DDR2-667 2.2V 128Meg x 64
(50%)2GB kit (1GBx2) BL2KIT12864AA1005 DDR2 PC2-8000 5-5-5-15 UNBUFFERED NON-ECC DDR2-1000 2.2V 128Meg x 64
(50%)2GB kit (1GBx2) BL2KIT12864AA804 DDR2 PC2-6400 4-4-4-12 UNBUFFERED NON-ECC DDR2-800 2.2V 128Meg x 64
(50%)1GB kit (512MBx2) BL2KIT6464AA1005 DDR2 PC2-8000 5-5-5-15 UNBUFFERED NON-ECC DDR2-1000 2.2V 64Meg x 64
(44%)512MB kit (256MBx2) BL2KIT3264AA53V DDR2 PC2-4200 3-3-3-10 UNBUFFERED NON-ECC DDR2-533 1.9V 32Meg x 64
(41%)512MB kit (256MBx2) BL2KIT3264AA664 DDR2 PC2-5300 4-4-4-10 UNBUFFERED NON-ECC DDR2-667 1.9V 32Meg x 64
(41%)2GB kit (1GBx2) BL2KIT12864AL664 DDR2 PC2-5300 4-4-4-10 UNBUFFERED NON-ECC DDR2-667 1.9V 128Meg x 64
(33%)4GB kit (2GBx2) CT2KIT25664AA53E DDR2 PC2-4200 CL=4 UNBUFFERED NON-ECC DDR2-533 1.8V 256Meg x 64
(33%)2GB kit (1GBx2) CT2KIT12872AA53E DDR2 PC2-4200 CL=4 UNBUFFERED ECC DDR2-533 1.8V 128Meg x 72

i have ordered these by the % score and as you can see that the voltages are also ordered, perhaps there are different chips being used, or the manufacturers are using the same chip for all of the above and using different voltages to gain performance, I do not know. If you have bought high performance ram and want to overclock and find that it aint got much, maybe the manufacturer beat you to it! (I wonder if this applies to CPUs too).

I was never able to test it performance-wise until recently, as I didn’t have a collection of sticks to test with. I just upgraded a PC from 33.333% to ram with a 50% rating (though I think it was running at 44% because the board’s AUTO setting due to SPD). Therefore I expected a bandwidth increase of +50% (or +33% if the ram was only running at 44%).
i will elaborate on my math because the figures will confuse.
33% ram does not mean +33% performance increase.
i have a 33.3% stick. and i have a 50% stick. the 50% stick is 50/33.3 or +50% faster (total 150%).
But I only got a bandwidth increase of +6.3%. Adding a second bank of 33% to the first 33% increased the bandwidth by +5.2% but this was because the stock sticks “featured” C6 At 400 in the SPD, and the antique sticks I added had C5 at 333 as the highest frequency. So the board chose C5 and sped up (i dont know if the voltage increased/should be increased any).
Increasing the FSB gave me a simular percentage memory bandwidth increase, while overclocking the CPU (I tested in case I got a performance hike when the cpu speed increased). So it could be the Ram frequency (or the FSB) that is the limiter. The “PC2” rating means little when comparing the chip latencies, and may instead allude that the DDR frequency is the bandwidth limiter.

The biggest letdown is that some overclocking motherboards lack the ability to fully configure the ram clock. If I am able to specify double the clock speed of the ram I could also double the latencies as well and then finetune with part of an integer (as if i were at the former frequency). Some boards will allow you to select a ratio, but with some the ratio choices are limited which can require larger increases to the frontside bus speed. (I did have a ratio board but its now not available to test on - bios).

it could be that the north bridge is the bottleneck as its not the ram or the CPU. So if anyone considers overclocking (or upgrading) the ram, you probably shouldn’t until you have overclocked the north bridge. And until then (should my suspicions about 33% ram be true), 33% ram is good enough, but it will need heatsinks when you do decide to overclock.

i am not sure if the bandwidth is as much of a problem with DDR3 but i guess i will find out if i ever decide to upgrade. (even though i didnt get much extra with DDR2 clock increases)

It could be that the reason why i got a small bandwidth increase when ram speeds increased, may have something to do with whether the ram nanoseconds and the board are in sync. The bandwidths may actually decrease for the same reason.

My theory is plausible but untested. It may require advanced tweaking of all the latency figures to get this going correct.

Following any recommendations from this text is entirely at your own risk and nothing is guarenteed. opinions from seasoned overclockers are appreciated.

Edited by Charlieb000 - 1/31/12 at 7:41pm
post #2 of 2
Thread Starter 
no comments. oh well maybe someone will find this useful in the future
Corrections: the board was set to "400" which is why it didnt go back to 33.3%, it went to the lower multiplier once it was returned to auto. again i saw the same speed differences, so it looks like its limited by the northbridge. so next would be lowering the CPU/RAM multipliers and raising the North (which raises the CPU and ram again) obviously an AMD only performance mod ;( meaning i am supposed to put it in the AMD memory section.

ValueOfRam.zip 91k .zip file
uses macros to expand the tables to make file smaller (was on dialup at the time)
Edited by Charlieb000 - 1/31/12 at 7:22pm
New Posts  All Forums:Forum Nav:
  Return Home
  Back to Forum: Memory
Overclock.net › Forums › Components › Memory › Where is the bottleneck in RAM access?