DDR2 vs DDR3 - ASUS Blitz Formula vs Blitz Extreme...CellShock Rams involved...

[hipro5]

lol....975x rules 3Dmark, no doubt.
395*10 cpu, 743/1000 2900xt 1gb*2 Crossfire.(no overclk)
ram=988 4-4-4-5, pcie=118mhz
mobo=Asus P5W64 ws pro
3Dmark05=264xx

Yes I know......After 2 years now or so, 975 STILL shows it's teeth..... Maybe because, as I wrote above, is some kind of latency involving of the "Crosslinx" "chip" is used to "break" the 1 x16 to 2 x8.....

[jugeen]

nice job!

[JMKS]

So... it seems that i965 with that 514,6 4:5 if it will be possible by both chipset [1066 strap... - with modded strap {which can be if correctly modded, strange, faster than genuine 1066 and less NB-stressing} it should be possible with subzero NB cooling I think] and memory [642 4-4-4-4] will be faster than P35 DDR3 1:2 8-6-6-x, that's also a bit strange .

My own tests:
DDR2 P35 vs. i965 using 16M Pi, all with 400x8 = 3,2GHz Allendale with various dividers and timings:


P35 5:6 [used by hipro5 also, maybe a bit looser timings] 4-4-4-4 2t 7.03.641s


P35 4:5 4-4-4-4 7.01.391s


i965 4:5 4-4-4-4 6.56.328s
Over 5s better than P35 with 4:5, so over 7s better than 5:6 with P35 - and it's 16M so in 32M the difference will be more than 15s surely... Rescaling it to 3,6GHz the difference "P35 5:6 vs. i965 4:5" will be still much more than 10s .

[hipro5]

Re-run your benchies with the P35 chipset coz it's "tricky" as for it's straps..... I bet you could do better with it.....

For example: The Extreme (DDR3) mobo is WAY FASTER with the 266 strap instead of the 200 which suppose to be faster......NO it's NOT....The 200 strap acts like the 333 one.....

I ALSO see that you don't have the same latencies as for your 4:5 benchies in 965 and P35......Please keep them EQUAL so to equal compare between them....For example: Write to Precharge Delayed make it 9 instead of 10 in the P35 and ONLY from this one, you'll get another 1 - 2 sec....

EDIT: Hmmmm....... I see another 1 - 2 sec lower in P35 by changing the tRTP from 6 to 4......

Ahhhh........You're cheating the P35 chipset......

ALSO keep in mind that we are talking of ~10sec LOWER in 32M at the EXACT SAME config....We "fight" at 3600MHz for .5 - 1sec lower score and we get a 10sec difference just like that from DDR2 to DDR3.......
I have used EVEN the SAME HD as for the DDR2/DDR3 and 975 chipset tests.....

[Johnny Bravo]

George ygpm

[JMKS]

Thanks for the lesson .

I will probably jump straight to DDR3 with P35 if I will have chance .

In that tests I wanted to max out both platforms so Performance Level and subtimings was as low as the chipset could do [from short tests], so not equal for both chipsets.
I don't remember to gain something big from mentioned 10 => 9 Write to Precharge Delayed or 6 => 4 tRTP but maybe I overlooked something or board was unstable with that settings - I really don't remember and I don't have P35 at this moment to re-run this.

[hipro5]

OK.......Let's some up.....

ABIT IN9-MAX with i680 chipset......and please don't tell that these Ram timings aren't good......



ASUS P5B-Deluxe with i965 chipset.....



ASUS Blitz Formula with P35 chipset AND DDR2 Rams.....



ASUS Blitz Extreme with P35 chipset AND DDR3 Rams....




Hmmmm.....now that I see ALL of them together, some thoughts came into my mind......
How was it possible for "some guys" to get Super-Pi 32M World Records with that crappy i680 chipset?......

[Xenogias]

nice summation Hipro. Now that you placed all teh chipset scores together and listed some 3d score I'm willing to give DDR3 another chance. I suppose its not bad for early stuff though I honestly wanted more out of a new ram standard. I'm thinking with x38 and 2nd generation DDR3 (higher clocking) we'd be talking about 12m 30s spi32M at 3600mhz...How does it scale with a higher clocked cpu?

[kiwi]

I always love to see sub 13m @ 3.6ghz

[hipro5]

nice summation Hipro. Now that you placed all teh chipset scores together and listed some 3d score I'm willing to give DDR3 another chance. I suppose its not bad for early stuff though I honestly wanted more out of a new ram standard. I'm thinking with x38 and 2nd generation DDR3 (higher clocking) we'd be talking about 12m 30s spi32M at 3600mhz...How does it scale with a higher clocked cpu?

If you're going to do 3Ds, I will suggest you to go with the x38 chipset and DDR3....

At 3950MHz it scales pretty well to..... ....AT 494MHz fsb x 8 and the Rams 1:2 and 988MHz with Cas 7 - 6 - 5 - 17 or EVEN 8 - 6 - 5 - 17, it does very well.....

I always love to see sub 13m @ 3.6ghz

Me too......

[filmbot]

Hipro, just for comparison sake, can you edit in 975 times also?

And I've always figured that DDR3 would take over DDR2 just for the sake of the raw speed. Plus you have more FSB on DDR3. No, offense to you, but I think in general you cant really straight compare DDR3 with DDR2 because of straps, dividers, FSB differences.

But great compaison with 680i, 965 and P35 (DDR2)

[hipro5]

Hipro, just for comparison sake, can you edit in 975 times also?

And I've always figured that DDR3 would take over DDR2 just for the sake of the raw speed. Plus you have more FSB on DDR3. No, offense to you, but I think in general you cant really straight compare DDR3 with DDR2 because of straps, dividers, FSB differences.

But great compaison with 680i, 965 and P35 (DDR2)

Hmmm.......Now that you have mention it, I don't believe that I have a 32M Super-Pi run with the i975......I'll try and see if I find anything in my HDs and if not, I'll give it a try.....

As for the others, yes you can't compare all of them "equally" BUT I always try to get the MAX out of any mobo and then compare it with a MAX of another mobo at the SAME CPU frequencies (for reference)..... I'm not a reviewer who just plugs in and runs.....

[dinos22]

[B]
As for the others, yes you can't compare all of them "equally" BUT I always try to get the MAX out of any mobo and then compare it with a MAX of another mobo at the SAME CPU frequencies (for refference)..... I'm not a reviewer who just plugs in and runs.....

[v_parrello]

Hipro5 very useful! Thanks a lot!

I understood at yee moment 975 is the best performer in 3D with crossfire.

What about 975 in superPI 32M. It performs better than P35? What is your opinion? Do you have some screenshot for 975?

Thanks a lot.

[saaya]

wow george, wow!
very nice! 2.4v... those poor poor chips
ok... i admit it, if id have a blitz id have tried 2.4v and higher too ^^

I did a Super-Pi 32M WITHOUT MAXMEM here with Cas 7 - 6 - 5 - 17 and 2.42Vdimm BUT after that, my ram modules never worked over 1000MHz at that Cas latencies...... Though they don't produce errors lower or with Cas 8.....
IF this was a bench with MAXMEM involved, it would be around 12.43.xxx.......

so even when the mem is cool it gets damaged at 2.4v.... good to know!
even with our thick aluminum heatspreaders pressed against the chips with thermal paste... but you said they were cool to touch... so it seems heat was not the problem at all hmmmm

did you notice any other degradation of the memory?
please check if it gets worse and let us know!
so atm what would you say is safe?
2.35v? 2.3v?

Yes I thought so (by thinking of A64's 1T vs 2T).... BUT NO......Slightly better.....Though wherever my rams overclocked at 2T, they did the same with 1T also.....So I prefaired the 1T....

but you needed more vmch to get the same speed 1T right? or did you straight go for a high chipset voltage and didnt check how high you can get with a certain chipset voltage?

I think that ASUS Blitz Extreme IS stable enough with them......It plays EASY with ALL the dividers and you can see the differences of them......
I do the WinRAR bench on every changing of divider when booting into windows to see how it does......It's fast, easy and you can see the "first" deferencies between Straps/dividers/etc.....Then you bench other proggies to test....

winrar, nice tip!
thanks

your posted loads of infos man, too much! hahaha
thanks for sharing all those results and tweaks!

[dinos22]

electron migration?!?

doesn't mean the sticks have to get hot to get damaged but if DDR2 is anything to go by 2.1v is where the critical limit is for DDR3 according to reports Tony had a while back with DDR2 (~39% over stock vdimm)

[Johnny Bravo]

electron migration?!?

damnit you beat me to it dinos!

[dinos22]

damnit you beat me to it dinos!

i've seen what EM does to DDR2 sticks >> actually the kit i have in this 24/7 system suffers from symptoms

i cannot boot at stocks settings and have to raise vdimm to 2V to get the rig to boot (lucky 256MB D9GKW jump sticks come in handy heh) but it still runs my usual OCs at higher vdimm >>>> funny thing is that these sticks haven't really seen much past 2.2v but who knows maybe P5B DLX spikes as well

i haven't killed any of the PC10K dominator kits i have even though i mainly run them at 2.4-2.45v >> fingers crossed it stays that way even thought i will give them hell with the X38 DDR2 mobo arriving early next week

[Gautam]

Example:

3DMark2005 at 3950MHz and CrossFire at 850/950

ASUS Blitz Formula: ~25600marks
ASUS Blitz Extreme: ~25994marks
ABIT AW9D-MAX: ~26422marks

2k3 with the 975 was ~73200marks

975 chipset was running at only 439MHz fsb, rams 1:1 with Cas 3-3-3-4....

As I see, 975 STILL rules in 3D till X38 comes out.....

For me, 437.5x12 3-3-3 on 975x was about 150 pts slower in 05 than 525x10 5:6 4-5-4 on Blitz. Not a huge difference but the P35 appeared a little faster.

[over@locker886]

Thanks, you share your tests/+experiences! Very useful..

[PanteraGSTK]

Wow, you were busy while I was sleeping. Interested in seeing the 06' benches. Hopefully x38 will be on par with 975. We will see soon enough...

[hipro5]

Hipro5 very useful! Thanks a lot!

I understood at yee moment 975 is the best performer in 3D with crossfire.

What about 975 in superPI 32M. It performs better than P35? What is your opinion? Do you have some screenshot for 975?

Thanks a lot.

You're welcome....
No I don't have 32M of the 975.... BUT I'm sure that it'll perform worse coz
a. It will not have high fsb and
b. It will not have high memory bandwidth....
Though it has lesser latencies than the i965, i680 and P35 which counts most on 3Ds....
If we could have ALL of them in one chip, then it whould be a blessing.... High FSB - HIGH memory bandwidth and LOW latency....

wow george, wow!
very nice! 2.4v... those poor poor chips
ok... i admit it, if id have a blitz id have tried 2.4v and higher too ^^

so even when the mem is cool it gets damaged at 2.4v.... good to know!
even with our thick aluminum heatspreaders pressed against the chips with thermal paste... but you said they were cool to touch... so it seems heat was not the problem at all hmmmm

It doesn't matter if one chip is cool or not when we talk about HIGH voltage/current.....EVEN 1sec of heavy load could damage it no matter what....In 1sec it will not even get hot....

did you notice any other degradation of the memory?
please check if it gets worse and let us know!
so atm what would you say is safe?
2.35v? 2.3v?

You can't say what is safe and what not in such occasions and here's why:
You may feed it with - let's say - 2.3Vdimm and you're "pussing" it too hard OR it has a "by manufacturing problem" (one chip at a million/billion)....IF ONE internal Mos-Fet of the entire SINGLE chip gets damaged, then the whole ram module IS damaged.....and we are talking about hundrents/thousands/millions Mos-Fets into ONE chip.....and Ram modules have 8 or 16 of them....

but you needed more vmch to get the same speed 1T right? or did you straight go for a high chipset voltage and didnt check how high you can get with a certain chipset voltage?

GOOD question....I had the same Voltage for 1T or 2T.......BUT we have a problem.....and HERE IS THE PROBLEM.....and a BIG ONE....
Feeding a P35 Northbridge chipset with HIGH Voltage IT GETS DAMAGED......YES YOU READ IT WELL.....IT GETS DAMAGED.....
I used to play with 1.74 to 1.81Volts on this mobo and I was OK....BUT after some crushings on 3Ds, it NEEDS 2.01Volts now to work my rams high....
First I though that was the rams produsing errors and crushings so I fed them with 2.34, 2.36, 2.38, 2.40, 2.42, 2.45Vdimm with NO LUCK at all.....
I then fed them with 2.34Vdimm and UPED the P35s Voltage to 1.91V.....I then was OK.....BUT NOT FOR LONG.... Keep upping the P35s voltage so to stabilize my rams, now I NEED 2.01Volts to do so.....DUMN this is NOT a good sign at all.....and we are talkin about 2 - 3 days period of time....

winrar, nice tip!
thanks

your posted loads of infos man, too much! hahaha
thanks for sharing all those results and tweaks!

You're welcome....

For me, 437.5x12 3-3-3 on 975x was about 150 pts slower in 05 than 525x10 5:6 4-5-4 on Blitz. Not a huge difference but the P35 appeared a little faster.

Maybe due to low fsb on 975 and high fsb on P35.....
As I figured out with some older testings of mine on the 975 goes something like this:
FSB 400MHz with Fatbodies at Cas 3-2-2-1 equals with FSB 440MHz with GKXs and Cas 3-3-3-1.....So at only 437 it's a bit semi-low.....At 450MHz with Cas 3-3-3-1, GKXs starting to show their teeth on the 975.....

Wow, you were busy while I was sleeping. Interested in seeing the 06' benches. Hopefully x38 will be on par with 975. We will see soon enough...

I'll put the QX6700 on the mobos to test it....I'm curius too you see....


.....and now some other testings just "for the eye".....Actually I'm not fond of them but I'll upload them for someone who might be interested about them.....

EVEREST Memory Bandwidth with Cas 8 - 6 - 6 - 18....



EVEREST Memory Latency with Cas 8 - 6 - 6 - 18....



EVEREST Memory Bandwidth at a bit higher frequency and Cas 7 - 6 - 5 - 18....



.....and the same with the Memset....



Some S-Sandra now........Actually it reads that I have 12GB of ram..... I don't like Sandra any more coz it aint accurate and it writes whatever it wants......BUT I'll upload it coz......it is full of colours.....

[Counter CS]

Very nice thread, we have a lot of information .

Subscribed

[PanteraGSTK]

Some S-Sandra now........Actually it reads that I have 12GB of ram..... I don't like Sandra any more coz it aint accurate and it writes whatever it wants......BUT I'll upload it coz......it is full of colours.....

Full of colors

[The Stilt]

Nice job there George

Since my board can´t do over 1900MHz DRAM CLK in dual channel I tried single channel instead.

I wonder how high the Everest score would be with the same clocks at dual channel.

Few tweaks included

SINGLE CHANNEL