Phenom II ddr III speed vs timings

I think the screenshots speak for themselves........Lets see some other comparisons of other benches to get a bigger picture. I suggest memory intensive...I will recompare 32m later but didn't have 1 1/2 hours to kill....

Method used, opened far right PI set affinity to core 1, priority high ran 16k twice then blasted a 1m, opened another PI 1m set affinity to core 2, priority high etc etc you get the idea......

http://members.cox.net/wmdieselmc26/...33%206-6-6.JPG

http://members.cox.net/wmdieselmc26/...00%207-7-7.JPG

lower speeds and tighter timings seems better

Maybe you could try runs that last about 5 mins each instead of the 32M or of course I can do it too instead of having you all the work. :-)

Quote:

---

Originally Posted by shaolin95

Maybe you could try runs that last about 5 mins each instead of the 32M or of course I can do it too instead of having you all the work. :-)

---

Sure have a shot at it, but if the times are a 10th of a second off now longer benches it will be like 1 sec off.......even so it just makes more sense to run tighter and slower as that 1 second might make a 1/100th of a FPS difference......but its alot easier to run tight and low ;)

I am actually using running tight and low myself right now but I am always curious. :up:

So when I was looking at that $60 OCZ D9 kit that runs DDR3-1333 @ CAS6 instead of DDR3-1600 at CAS8 then I was actually being smart instead of just cheap. Cool :cool:
Not like I'm buying soon anyway but good to know I was headed in the right direction.

on ddr2 didnt speed matter more than timings?

Sure but ddr III is a different animal now,+ with ddr III I think we have reached the ceiling and we have a new revision of the IMC.

so lets say i was going to get the gd70 w/ a 955be and 8gb ram would you get:

1600mhz 7.6.6.24 1.9v kit
or
2000mhz 9.9.9.28 1.9v kit

they are both ocz kits, do you think they could run those tight timings at lower mhz? i know w/ ddr2 sometimes faster ram couldnt get the tight timings at lower speeds (think 1000mhz 5.5.5.15 kits that cant do 800mhz 4.4.4.12)

is ddr3 the same way? i was leaning toward the 2000mhz kit just figuring having the headroom cant hurt



edit to add: i wonder how your results would scale w/ more nb freq and even cpu

have any other tests to show this? maybe the speed is already high enough on the ram that it would benefit from tighter timings.

Quote:

---

Originally Posted by crazydiamond

so lets say i was going to get the gd70 w/ a 955be and 8gb ram would you get:

1600mhz 7.6.6.24 1.9v kit
or
2000mhz 9.9.9.28 1.9v kit

they are both ocz kits, do you think they could run those tight timings at lower mhz? i know w/ ddr2 sometimes faster ram couldnt get the tight timings at lower speeds (think 1000mhz 5.5.5.15 kits that cant do 800mhz 4.4.4.12)

is ddr3 the same way? i was leaning toward the 2000mhz kit just figuring having the headroom cant hurt



edit to add: i wonder how your results would scale w/ more nb freq and even cpu

---

Thats a good point and yes it does scale but not at speeds anyone can reach with air or water.......seems the IMC is partial to 5-5-5 and 6-6-6 latencies.

I would suggest this stuff I'm playing with. Does 6-6-6 1333+ and 1600 7-7-7 at 1.65v for both......Sure there is faster stuff but at the price point its hard to beat.....Its what I'm using in the above posts.....

http://www.mwave.com/mwave/viewspec_...iteria=BA26056

I think optimum would be ram that can do 5-5-5 1333+ and 6-6-6 1600....gives you flexbility to tune it whichever way works best for your particular PC.

Will be testing other stuff over the next few days, I think i'm going to test fear and crysis to get some more realworld comparisons.

these are still early chip they have to mature for good memory support on ddr3.

these have to do DDR2 and DDR3 it's not surprising they can't reach above 1800mhz.

sure a bit disappointing, but not really a big deal.

later chips will probably get better as time go's on.

might be the reason why the imc on the cpus are just rated for 1333. later on the imc will become more powerful tho and im sure speed will be better than timings the same way it was for ddr2.

depends on whether the application needs fast memory access or high bandwidth.

i tested 1333 cl6 vs. 1600 cl7 when gaming and there was a remarkable difference towards 1600 cl7 was clearly better in a very problematic crowded area (MMORPG). I cannot prove that, but perhaps gaming benches would show (i dont have any recent games here though)

plus, super pi is not really memory intensive. its not tenth but some 1/100 difference there...

Quote:

---

Originally Posted by Oese

depends on whether the application needs fast memory access or high bandwidth.

i tested 1333 cl6 vs. 1600 cl7 when gaming and there was a remarkable difference towards 1600 cl7 was clearly better in a very problematic crowded area (MMORPG). I cannot prove that, but perhaps gaming benches would show (i dont have any recent games here though)

plus, super pi is not really memory intensive. its not tenth but some 1/100 difference there...

---

Check out these: ExtremeTech

yeah.. but i think ddr3-1333 timings are not really optimized there..

whatever, impressive numbers from the amd ddr3 system..

Is it stated anywhere on that article what timings were used for each? I cant see it specified and if its not then its such a pointless test.

PS I just got a reply from the author and he thinks he used 7-9-9-22 for both. Doesnt really help us much here as obviously the 1600 will be better if the timings are kept the same.

thats right then really.. thats the cause for 1333 was slower then ddr2 at higher clock then too..

but, what's nice with this is: even with ddr3-1600 with 7-9-9-22 (what is rather slow) the numbers shown are very nice for just 3.12ghz clock :)

What would be cool ,,

a Single IMC(onchip) that did both DDR2 and DDR3 thus use the Same DDR2 Dimm slots for DDR3. Not two seperate IMC's..Just One

Why didnt Mobo makers,,Make Bios or mobos to run Either DDr2 or 3 on the fly in the Same Dimm Slots.

Two sepperate IMC's are being used ATM instead of a single IMC,, thus you must have a DDR2 or DDR3 Mobo not Both on the Same Mobo .

It would be Nice to upgrade from DDR2 to DDR 3 on the Same Mobo as how PHII can run AM2+/AM3.

Its odd why DDr3 cant use a DDr2 slot (IMC/Bios is what controls the Ram ) even though it has same amount of pins/tabs you'd think they would of made DDR3 compatable to drop into a DDr2 slot Via a Bios or something.

From my understanding the Slots/Rams differ only by Volts and how there pinned with the exception of being faster chips . DDR2/DDR3 side by side look the same besides speed/latency.

I always thought DDR3 was just a DDR2 PCB with improved chips. Kinda like what they did with Vidcards. You know when they went from DDR2 to DDR3 on the Same card. Manufactures used the Same PCB's for DDR2/DDR3 videocards.

Having the Same Memory PCb DDR3 to drop in to a DDR2 slot they would only need a AM3 IMC that could run all DDR2/DDR3 dimms instead of two seperate IMC's for DDR2/DDr3 .


Latencies would be a lot better i would think.

Hahaa that Was my two bits worth:D

In the end ive always noticed the PH1/PH!! runs best with a medium Latency 5/6/7 and High NB above 2400mhz with Medium Ram Freqency say 1100+ to 1200mhz.

Untill we see a Single AM3 IMC DDR3 only setup (no DDr2IMC involved) then well see Way faster Numbers @1600+mhz and even higher.So in the end i think Higher ram with 5/6 timing will be seen even @ 2000mhz eventually.

ddr3 has the center groove in a different slot, it wont fit into ddr2 slots

Quote:

---

Originally Posted by crazydiamond

ddr3 has the center groove in a different slot, it wont fit into ddr2 slots

---

You made my headache go away :ROTF: That explains it,, so a whole new ball game with DDr3 then...Not as simple as i thought between the Two types of Dimms.


So AMD needs to unleassh the Pure AM3/DDR3 only chip (no ddr2IMC),,shure to be " The InHell Inside killer:explode2:":D

why does it matter if it has 2 imcs? i dont think you would get any performance loss it just costs slightly more to make the chip.

Quote:

---

Originally Posted by roofsniper

why does it matter if it has 2 imcs? i dont think you would get any performance loss it just costs slightly more to make the chip.

---

because it's added transistors I believe phenom II has 758 million transistors
vs the 731 million in I7, the extra transistors are for ddr2 and and ddr3 the difference being that ddr3 uses a 4 bit wide and ddr2 uses 2 bit wide addresses.
so you would need two different IMC's.

i know but even if i was going pure ddr3 and never going back i really dont see why it matters if it has 2 imcs. i know why they are both there but sure take out the ddr2 controller and thats a few less transistors. i think if you are running ddr3 anyway the ddr2 controller isnt even burning any power so really it doesnt seem like its that big of a deal. besides we all know that soon enough amd is having certain cpus only ddr3. i believe istanbul is ddr3 only. deneb only had both to offer a smooth transition.

Quote:

---

Originally Posted by roofsniper

i know but even if i was going pure ddr3 and never going back i really dont see why it matters if it has 2 imcs. i know why they are both there but sure take out the ddr2 controller and thats a few less transistors. i think if you are running ddr3 anyway the ddr2 controller isnt even burning any power so really it doesnt seem like its that big of a deal. besides we all know that soon enough amd is having certain cpus only ddr3. i believe istanbul is ddr3 only. deneb only had both to offer a smooth transition.

---

guess it comes down to price. :shrug:

I have been testing my Kingston Hyper-X 1800 modules and so far have 1.7v and cas 7 @ 1600mhz. What do you think Chew*? I think it is good settings, I usually have my NB @ 2400 and CPU higher, but wanted to make sure any errors were infact the memory.

http://3800z24.info/Phenom/x3_720/1600-cas7-1.7v.jpg
http://3800z24.info/Phenom/x3_720/1600-cas7-1.7v_2.jpg

Remember phenom am2 had a 1066 mode...things slowed down to do 1066 and you had to force NB clock higher to claw back performance. Am3 has the same issue when you go to 1600, the only issue is you need to jump to 3000MHZ on the NB to gain the performance back, 2800 and 2600 show no difference from my testing here.

So 1333 6-6-6-24 is the sweet spot, 2600MHZ NB and 2600HT, easy clock with 1.6 to 1.65V ram and the CPU will probably go a few MHZ higher also due to less strain.

You don't need high ram with AM3...for 24/7 its not worth the effort

Quote:

---

Originally Posted by Tony

Remember phenom am2 had a 1066 mode...things slowed down to do 1066 and you had to force NB clock higher to claw back performance. Am3 has the same issue when you go to 1600, the only issue is you need to jump to 3000MHZ on the NB to gain the performance back, 2800 and 2600 show no difference from my testing here.

So 1333 6-6-6-24 is the seet spot, 2600MHZ NB and 2600HT, easy clock with 1.6 to 1.65V ram and the CPU will probably go a few MHZ higher also due to less strain.

You don't need high ram with AM3...for 24/7 its not worth the effort

---

So I should try for 1333 and even tighter timings? I think I can get below 1.7v at that speed and CAS 6, what do you think?

I totally agree with Tony as I was playing with my RAM today and found out exactly what he is stating her abour the NB at 2800 not doing much when running 1600.

Quote:

---

Originally Posted by shaolin95

I totally agree with Tony as I was playing with my RAM today and found out exactly what he is stating her abour the NB at 2800 not doing much when running 1600.

---

I am going to try and recreate this with wprime, 3dmark CPU score, WinRAR. with lower RAM and tighter timings. You got me currious.

I also used pcmark, vantage and crysis CPU test.

ehhh im still buying 1600 ram but i might just downclock it if its better.

I will run more test after wife goes to bed, well maybe:p:

Here is what I did with 2000NB. the CPU is clocked 3.4 do to my AOD profile loading with wPrime ( I was testing rpofiles and have changed it). what I noticed was wPrime actually looked to do better, Pi was about the same, and WinRAR fell some. I will raise the NB higher and run test again.

First is 1333 6.6.6 21

http://3800z24.info/Phenom/x3_720/13...6-3.4-test.jpg

Second is 1600 7.7.7 23

http://3800z24.info/Phenom/x3_720/18...7-3.4-test.jpg

im curious how this affects real world performance like games and other programs for example. seems like winrar would be a better test than pi or prime would be. its interesting to see this tho. with ddr2 there was no wall and the highest frequency you could get was always the best performance. now with ddr3 theres a wall right in the middle where if you go too high you get lower performance. wonder if as the imc on the cpu advances if this will change and end up like ddr2. what happens when you run 1333 at c7? similar to 1600 at c7?

i just see this as the nb is bottlenecking the ram and w/ higher nb speeds it would show improvements for higher speed ram... hopefully between bios and imc improvements we will start to see 3000mhz+ nb on air regularly :shrug:

interesting point indeed...I cant achieve 3000 unfortunately to test this.

Quote:

---

Originally Posted by shaolin95

interesting point indeed...I cant achieve 3000 unfortunately to test this.

---

I can only hit 3000nb on my 940. Both 720's have a 2800NB max and 2600 is most stable. I am going to keep my 1600 CAS 7 @ 1.7v. It has the best overall performance so far regaurdless of test.

Quote:

---

Originally Posted by shaolin95

interesting point indeed...I cant achieve 3000 unfortunately to test this.

---

I will be testing 3000 + nb in the next few days...........but as I stated before for those on air/water its pointless to get High speed ram unless you have plans to go sub zero........

Quote:

---

Originally Posted by charged3800z24

I can only hit 3000nb on my 940. Both 720's have a 2800NB max and 2600 is most stable. I am going to keep my 1600 CAS 7 @ 1.7v. It has the best overall performance so far regaurdless of test.

---

at what voltage ? mine needs to much like 1.625 volts and I've only done it once.

Highest I have done on air was 3200 IIRC, on the DS4H and I pushed it hard very hard to do so.....

Quote:

---

Originally Posted by demonkevy666

at what voltage ? mine needs to much like 1.625 volts and I've only done it once.

---

the 720's need 1.45 for 2800. it took the same on the 940 to hit 3000nb. The 940 is on the asus M3A79-t the 720's are both on MA790FXT-UD5P boards. I used f3L bios. I am back on f3B for some 4core loving. 2600NB is all I can get right now.

Quote:

---

Originally Posted by chew*

Highest I have done on air was 3200 IIRC, on the DS4H and I pushed it hard very hard to do so.....

---

that is pretty impressive. I never tried for more then 3000NB.

The gigabyte needs the NB fixed atm for 955's, thats why you see me running 2600 on it......taking +.300 to do that and it shouldn't, the msi is doing 2800 right now at 1.32 for comparison.........

Quote:

---

Originally Posted by chew*

The gigabyte needs the NB fixed atm for 955's, thats why you see me running 2600 on it......taking +.300 to do that and it shouldn't, the msi is doing 2800 right now at 1.32 for comparison.........

---

wow, I need a 955, I wonder when that will be fixed on the Gigabyte? I don't really wanna buy another board:p:

Quote:

---

Originally Posted by charged3800z24

the 720's need 1.45 for 2800. it took the same on the 940 to hit 3000nb. The 940 is on the asus M3A79-t the 720's are both on MA790FXT-UD5P boards. I used f3L bios. I am back on f3B for some 4core loving. 2600NB is all I can get right now.

---

was that in XP, cause I'm vista 64 bit remember.

too bad gotvoltlage isn't in here with his results form his 9950 from 2700mhz NB and higher and low memory.

Quote:

---

Originally Posted by charged3800z24

wow, I need a 955, I wonder when that will be fixed on the Gigabyte? I don't really wanna buy another board:p:

---

Very soon I hope, If not I will give them a nudge and show screenies and voltages required to do x and the same ram at higher clocks on another board to show them the light........

Yes I prefer to use them but I'll be the first to call them on an issue as well, a few betas work fine......I want the public to work fine........

Fanboys don't get anything fixed, dirgruntled end users do.

Quote:

---

Originally Posted by demonkevy666

was that in XP, cause I'm vista 64 bit remember.

too bad gotvoltlage isn't in here with his results form his 9950 from 2700mhz NB and higher and low memory.

---

win xp 32 and win 7 32.

well see how it works out with a high nb. the ram controller is speced for 21 GB/s and 2 sticks of 1333 is slightly over 21 GB/s. 2 sticks of 1600 would be 25.6 GB/s. dont know if i am calculating it right but to get full performance out of 2 sticks of 1600 you would need the nb to run at about 2.7ghz. try something like that and see how it performs then.

Quote:

---

Originally Posted by roofsniper

well see how it works out with a high nb. the ram controller is speced for 21 GB/s and 2 sticks of 1333 is slightly over 21 GB/s. 2 sticks of 1600 would be 25.6 GB/s. dont know if i am calculating it right but to get full performance out of 2 sticks of 1600 you would need the nb to run at about 2.7ghz. try something like that and see how it performs then.

---

this reminds me I once said your memory speed is best at half of your Nb speed. and that's extremely close, even if is rounded off.
which would mean you would need 3.6ghz NB for 1800mhz memory.

Quote:

---

Originally Posted by roofsniper

well see how it works out with a high nb. the ram controller is speced for 21 GB/s and 2 sticks of 1333 is slightly over 21 GB/s. 2 sticks of 1600 would be 25.6 GB/s. dont know if i am calculating it right but to get full performance out of 2 sticks of 1600 you would need the nb to run at about 2.7ghz. try something like that and see how it performs then.

---

We've already tried it..........wouldn't post info like this if we hadn't.........that why I said 1600 is useless for air/water.......you need 3000nb + and cpu speeds in excess of 4+ as well....it scales linear yes, but not at speeds doable 24/7air/water.

At 6.2g 4000+ NB we tested 1800 7-7-7 versus 1700 9-9-9 11.078 PI versus 11.02x we lost screenie of the 11.02x...........the difference is very very slight..........really need 1800+ 6-6-6 to low volt stuff to scale.....with 4000 + NB.............

If you remember msimax compared 3 boards as well 2600 NB was faster with 1600 ram than 2800 was.......

yes ddr3 has brought on a new wave of confusion. ddr2 was pretty straight forward and you would go for the highest clocks you could get now its just like wtf guess check and revise? well my other calculation was around 3.2ghz nb which would make sense and also agree with demonkevy's though too. but wouldnt 1800 7-7-7 always be faster than 1700 9-9-9 because the timings are lower and the clocks are higher? if you got anything else to show that would be nice. it might take time to get used to this whole ddr3 thing. but hows it on intel's side? did they ever have similar issues on core 2 or i7 with ddr3? seems kinda pointless to be making ram that can run at speeds like 2200 mhz if speeds much lower seem to perform better.

yes 1800 7-7-7 should be always faster than 1700 9-9-9 but it was not faster by a large margin as was previously seen with ddr 2.......

.05 in PI 1m is small potatoes considering those timings and speeds. In real world we are talking about 1/10th of a FPS better in a game? not worth the effort or money ;)

well how can you explain this? the ram is too fast for the imc? the ram is too fast for the cpu itself? how can this be fixed? because really i see no point in ddr4 if they cant find a way within the next few years to take advantage of all the bandwidth. unless they make a gpgpu.

Quote:

---

Originally Posted by roofsniper

well how can you explain this? the ram is too fast for the imc? the ram is too fast for the cpu itself? how can this be fixed? because really i see no point in ddr4 if they cant find a way within the next few years to take advantage of all the bandwidth. unless they make a gpgpu.

---

they have either eliminated a bottleneck and or created another bottleneck........maybe something we won't be able to get around until we get a new NB or SB?

The other possibility is they tuned it that way for a reason.......so that air/water users would get the max performance without killing themselves to get high clocking ram stable.........It not like the platform is bandwidth starved atm anyway..........

Quote:

---

Originally Posted by chew*

maybe something we won't be able to get around until we get a new NB or SB?

---

this is what i was thinking between imc revisions and the new chipset coming out hopefully we can start to see some nb clocks that make the ram shine

ok did some testing with latency dependant super pi 1m (1600 mode vs. 1333 mode) to see if there's really a slowdown with 1600 mode:

http://www.abload.de/thumb/pi1333mode06eg.jpg http://www.abload.de/thumb/pi1600modem14u.jpg http://www.abload.de/thumb/pi1600modeslownbumch.jpg http://www.abload.de/thumb/pi1333modeslownbf7tm.jpg

to compare, i'll calculate the average of each of those two runs:

run 1 (1333 mode): 18.545s
run 2 (1600 mode): 18.565s
run 3 (1600mode slow nb): 18.617s
run 4 (1333mode slow nb): 18.626s

i think not much difference at all. variability in those runs is higher then difference in mean values. differences are most likely found to be due to nb speed... ~50ms count for 200mhz nb, so 10ms can easily be due to 50mhz more or less nb speed when comparing run 1 to run 2... thus no problem with running 1600 divider i think..

next up i'll do vantage and 3dm06 to compare 1333 cl6 with 1600 cl7 as a more bandwidth-dependent bench...

1M is not memory intensive as the large cache of the cpu is mainly used for 1M calculation instead of the ram, hence the faster 1M calculation time when raising NB-CPU frequency. (=faster acces to the cpu cache)

If you want to test memory performance 32M is best but it will take ages to test.

agree, thats why there's no difference between 1333 cl6 and 1600 cl7. bandwidth does not count here. latency is a factor though. one can prove that by comparing ddr3-1333 cl7 to ddr3-1333 cl6.

http://www.abload.de/thumb/pi133365518j3xb.jpg http://www.abload.de/thumb/pi133376618p4nz.jpg

thus, slower overall latency settings in the imc when 1600 mode is chosen should have an effect when doing a latency-dependent bench. Thats what i tested, and i didnt find such....

here you can see the effect of latency/bandwidth as well:

http://www.abload.de/thumb/1333666182m7c.jpghttp://www.abload.de/thumb/160076620ocv8.jpghttp://www.abload.de/thumb/133365518zi9o.jpg http://www.abload.de/thumb/16007772002wa.jpg

1333: 6-5-5-18: 3dm06 21265, cpu 5510, sm2 8480, sm3 10337; vantage 13742, cpu 13654, graphics 13771
1333: 6-6-6-18: 3dm06 21217, cpu 5548, sm2 8378, sm3 10346; vantage 13703, cpu 13715, graphics 13700
1600: 7-6-6-20: 3dm06 21285, cpu 5514, sm2 8481, sm3 10355; vantage 13620, cpu 13690, graphics 13596
1600: 7-7-7-20: 3dm06 21236, cpu 5482, sm2 8441, sm3 10372; vantage 13722, cpu 13638, graphics 13750

mmh not really consistent findings here... seems all is within range of variability..

Have found that 1700Mhz 6-6-6-15-18 is the fastest for me without a doubt! Bit of both high bandwidth and low latency!

haha lucky one ;)

Sorry about that it meat to say 1600Mhz 6-6-6-15-18
Need 2.0v for 1700Mhz 6-6-6-15-18 and am not happy using that voltage 24/7! Prefer using 1.65v for 1600Mhz 6-6-6-15-18

hehe sure ;)

Oh, good thread. I already had some thoughts regarding the need for high frequency DDR3 memory when comparing the AM3 with AM2+ motherboards.

http://www.madshrimps.be/?action=get...58&articID=926

Check out the Lavalys Everest results in the overclocked test setting: 750MHz 7-6-6-18 performs pretty much the same as 600MHz 4-5-5-16 :-)