OCing on NF4: Memory Bandwidth at diff. Freq. n Timings

T

TheMask

Explorer
I ran SiSoft Sandra 2005 Memory benchmarks st different frequency/timings and take screenshots of it. Since I cudnt run Prime 95 to check for stability, I ran Super_PI 1M benchmark. The lack of time and patience didn't let me run any other benchmarks.

Test rig:

A64 socket-939 3000+ (Winchester)

Asus A8N-E (nForce4 Ultra)

Kingston Hyper-X DDR500 2 x 256MB

First I set these values in the BIOS:

Advanced Menu > CPU Configuration > DRAM Configuration:

Memclock Index Value [400MHz]

CAS# Latency Tcl [2.5]

Min RAS# active time (Tras) [7T]

RAS# to CAS# Delay (Trcd) [4T]

Row Precharge Time (Trp) [3T]

1T/2T Memory Timing [1T]

S/W DRAM Over 4G Remapping [Disabled]

Advanced Menu > CPU Configuration > HyperTransport Frequency [4x]

Advanced Menu > CPU Configuration > Cool 'n' Quiet Control [Disabled]

Advanced Menu > Jumper Free COnfiguration:

Overclock Profile [Manual]

CPU Frequency [200.0]

PCI Express Clock [100MHz]

DDR Voltage [2.85V]

CPU Multiplier [x9]

CPU Voltage [1.4V]

PCI Clock Synchronization Model [33.33MHz]

Please note that for some tests I had set the Vdimm at 3V and Vcore at 1.6V.

P.S: The order in which the timings are posted henceforth is: CL-Trcd-Trp-Tras. All tests were run at 1T Timing.

The following tests were carried out with the HyperTransport Frequency (multipler) at 4, and Memclock Index Value - 400MHz, i.e., HTT and DRAM ratio at 1:1.

Test 1. HTT200/DDR400 CL2.5-4-3-7:

Int Buff'd - 5171MB/s

Float Buff'd - 5110MB/s





Test 2. HTT225/DDR450 CL2.5-4-3-7:

Int Buff'd - 5825MB/s

Float Buff'd - 5755MB/s





Test 3. HTT235/DDR470 CL2.5-4-3-7:

Int Buff'd - 6083MB/s

Float Buff'd - 6008MB/s





Test 4. HTT225/DDR450 CL3-4-4-7

Int Buff'd - 5776MB/s

Float Buff'd - 5707MB/s





Test 5. HTT240/DDR480 3-4-4-8

Int Buff'd - 6202MB/s

Float Buff'd - 6099MB/s





Test 6. HTT250/DDR500 3-4-4-8

Int Buff'd - 6421MB/s

Float Buff'd - 6346MB/s





Test 7. HTT260/DDR520 3-4-4-8

Int Buff'd - 6673MB/s

Float Buff'd - 6599MB/s



Notes from the above:

Super_PI failed to run the run the 1M benchmark at DDR470 2.5-4-3-7 and at DDR520 3-4-4-8, showing that the RAM is at its limit at those timings and frequency and the system is not stable. Any higher on the frequency i tried, I wud get a BSOD.

The following tests were carried out with the HyperTransport Frequency (multipler) at 4, and Memclock Index Value - 333MHz, i.e., using the memory divider.



Test 8. HTT200/DDR333 2.5-4-3-7

Int Buff'd - 4753MB/s

Float Buff'd - 4703MB/s





Test 9. HTT225/DDR370 2.5-4-3-7

Int Buff'd - 5341MB/s

Float Buff'd - 5285MB/s





Test 10. HTT235/DDR385 2.5-4-3-7

Int Buff'd - 5579MB/s

Float Buff'd - 5522MB/s





Test 11. HTT246/DDR400 2.5-4-3-7

Int Buff'd - 5837MB/s

Float Buff'd - 5770MB/s





Test 12. HTT275/DDR450 2.5-4-3-7

Int Buff'd - 6543MB/s

Float Buff'd - 6470MB/s





Test 13. HTT293/DDR480 3-4-4-8

Int Buff'd - 6889MB/s

Float Buff'd - 6820MB/s





Test 14. HTT305/DDR500 3-4-4-8

Int Buff'd - 7158MB/s

Float Buff'd - 7084MB/s





Test 15. HTT310/DDR508 3-4-4-8

Int Buff'd - 7284MB/s

Float Buff'd - 7210MB/s





Test 16. HTT280/DDR460 2.5-4-3-7

Int Buff'd - 6641MB/s

Float Buff'd - 6567MB/s



The following tests were carried out with the HyperTransport Frequency (multipler) at 3, and Memclock Index Value - 333MHz, i.e., using the memory divider.



Test 17. HTT280/DDR460 2.5-4-3-7

Int Buff'd - 6649MB/s

Float Buff'd - 6581MB/s



Tests 16 & 17 were run at the same HTT and DDR settings, except in Test 16 the HTT multiplier was set to 4x and in Test 17 the HTT multiplier

was set to 3x. Interestingly, the scores are even for both the tests.

From Tests 1 & 11 we see that though the RAM is running at the same frequency (DDR400) and timings (2.5-4-3-7), in Test 11 there is a significant boost to the memory bandwidth. This is b'coz of the higher HTT and use of memory divider in Test 11. The memory bandwidth is enhanced by a significant 12.8%. Similarly, comparing Tests 2 & 12 we see that using a higher HTT and memory divider boosts the memory bandwidth by 12.3%. Now, comparing Tests 6 & 14, the latter receives a boost of 11.5%. It is a known fact that on nForce 4 Ultra chipsets the memory bandwidth largely depends on the CPU clock speed too. So, it makes sense that even if u have not so OCing RAM, u need not worry. U cud crank up the HTT/CPU speed use a suitable divider and still get the most out of your RAM. Ofcourse, tighter timings make for better scores in benchmarks and in real world applications too. But my Hyper-X doesnt run on timings any 'tighter' than 2.5-4-3-7.

Also, when the RAM is at its limits though I cud run SiSoft Memory Benchmark, Super_PI didnt always run. I hope the info provided is of some use to you.
 
U r most welcome to do so mate! :) the more "reviews" we have, the merrier.. BTW, this is kinda reveiw of the Hyper-X.. except i didnt make it one :D
 
great job maan...... :ohyeah: :hap2:

nd tht wud be awesome Bottle..... :hap2: :cool2:
Hope to c it soon

DaMn ..if only v had more kinds of memory sticks here to chooose from.. :@
 
Atlast you completed the experiment masky. !!!!

Good work.
Bottle Kindly do your experiments too and add your readings to this database asap .....
perhaps both of you togather can devise a sort of performa / format so that many more members can post their results too in that format so that it is easy to co-relate the readings.
i would have loved to put my input too but since i currently do not have a AMD64 i am unable to do it......but soon ......;)
 
great work!!! Keep it up....

Wish I had some Gskill RAM or OCZ VX to upstage you all..... ;). But all I have now is Kingston VR.

And another thing I wanted to ask you guys..... The chips on my RAM reads like this --
Hynix 448AP
HY5DU56822BT ---- D43

Does this mean that my RAM is based on generic D43 Hynix chip?? Coz' if it is, then from Darky's oc'ing, I gather it is pretty decent RAM.
 
nikhilb2008 said:
And another thing I wanted to ask you guys..... The chips on my RAM reads like this --

Hynix 448AP

HY5DU56822BT ---- D43

Does this mean that my RAM is based on generic D43 Hynix chip?? Coz' if it is, then from Darky's oc'ing, I gather it is pretty decent RAM.
Click to expand...

Yes that is what it is. The thing is not all RAM (even if they have the same chips) OC to the same level. And generally 256MB sticks OC much better than 512MB sticks. Darky's high OCes were on 256MB Hynix D43 chips :)
 
Guys, it is not only the chips that determine the o/c ability of the ram.

The quality of the pcb plays a very important role in o/cing.
if the pcb is well designed and of good quality , meaning 6 layers, good grounding planes etc , which reduce noise.... then the o/c potential increases too.
hence you will find good quality rams have a better pcb then a similar generic ram. similar in the sense they use the same memory chips.
infact the hardcore oc'ers and knowledgeable ones at that first see the pcb of the rams. and by just looking at them they have a fair idea about its o/c ability (potentially...)
 
deejay said:
Guys, it is not only the chips that determine the o/c ability of the ram.
The quality of the pcb plays a very important role in o/cing.
if the pcb is well designed and of good quality , meaning 6 layers, good grounding planes etc , which reduce noise.... then the o/c potential increases too.
hence you will find good quality rams have a better pcb then a similar generic ram. similar in the sense they use the same memory chips.
infact the hardcore oc'ers and knowledgeable ones at that first see the pcb of the rams. and by just looking at them they have a fair idea about its o/c ability (potentially...)
Click to expand...

Heh ur joking right... there are only two types... JEDEC compliant and brainpower. AFAIK, the brainpower is cheaper and overclocks better. Hence used by most mem manufacturers. (edit: i mean overclocking memory manufacturers).
 
Thanks for the info.....

@Chaos. Please check out my overclocks at the "Post your overclocks" topic. And comment.
 
Last edited by a moderator:
Chaos said:
Heh ur joking right... there are only two types... JEDEC compliant and brainpower. AFAIK, the brainpower is cheaper and overclocks better. Hence used by most mem manufacturers. (edit: i mean overclocking memory manufacturers).
Click to expand...
No Chaos i am not joking , Should i be ....??:D
What you have mentioned is 100% correct:cool2:
but that was the point i was trying to make too. that the pcb plays a important role in o/cing and the quality of the pcb including the components used to manufacture them determines a "Great" ocing module from a "Good" one.
What you are mentioning is only the "type" of pcb. and i am also implying the quality of the pcb including the components used on them. for e.g. the type of resistors used, trace lenghts (it effects the capacitance of the circuit), and even the type and quality of solder used.

These make a difference in ocing however mimiscule they might be.
notice the difference in o/c levels between different co's modules although they use identical pcb's and identical chips.
BP modules are the current "choice" of the majority co's since they are a bit smaller in size than standard jedec ones. this means shorter trace lenghts and the location of resistors (bottom part of the pcb) which help in o/cing.also shorter in dimension means the cost is a bit lower to standard jedec ones.

but some of the modules using that pcb are priced very high.
Seaching google lead me to this page where interesting info is mentioned about co's using BP modules
but even within the BP pcb's range there are variations. there are many "revisions" of the same basic pcb layout. for e.g. B6U808 and B6U815
(a very long thread down below)
http://www.xtremesystems.org/forums/archive/index.php/t-43659.html
This page describes the selection of pcb on o/c performance

http://www.legitreviews.com/article.php?aid=53&pid=1
now even these BP ones are being replaced by newer versions of pcb's which should hit the market within the next few months promising higher clock rates. > 600 mhz DDR easily.
ps: by reading threads such as mentioned above i have found one more reason why co's use the so called heat spreaders. this is so that the consumer or end-user does not get a chance to actually see what lies underneath them.:ohyeah: unless they want to void the warranty.
 
deejay said:
by reading threads such as mentioned above i have found one more reason why co's use the so called heat spreaders. this is so that the consumer or end-user does not get a chance to actually see what lies underneath them.:ohyeah: unless they want to void the warranty.
Click to expand...

That is why Gskill is known as the best. :clap: No fancy heatspreaders for them. And Samsung TCCD chips do not produce enough heat even when overclocked to require heatspreaders. So, heatspreaders on fancy OCZ, Corsair RAM is a sort of waste.
 
nikhilb2008 said:
That is why Gskill is known as the best. :clap: No fancy heatspreaders for them. And Samsung TCCD chips do not produce enough heat even when overclocked to require heatspreaders. So, heatspreaders on fancy OCZ, Corsair RAM is a sort of waste.
Click to expand...
True but winbonds running 250+ MHz at 2-2-2 at 3.5V+ do require the heat spreader. In fact the new mushkin redline requires active cooling :P
 
Back
Top