So what really is faster...3.8GHZ CPU but NB at 2000 or 3000MHZ?

[lopri]

Updated to include 1600 results

Thank you much again, Tony. It's exactly what I wanted to see. I think I'm starting to get to know about PII. And probably you've known about it all along. :P My hypothesis regarding PII architecture from your tables above:

1. When data can be replenished from L3, K10's performance keeps increasing to NB frequency, until it reaches the same frequency as the core frequency.
2. When data needs to be fetched from system memory, increasing NB frequency alone will give diminishing returns unless memory can keep up with it. Ideally it should be [NB frequency = 2 x memory frequency] (dual-channel)

So in my imagination, a 'perfect' 955BE would be 3.2GHz core / 3.2 GHz NB / dual-channel DDR2-1600 or higher.

Does this make sense to you, or am I totally wrong?

[lopri]

I for one am simply curious at what speed for DDR2 (perhaps up to 600Mhz) the NB stops giving much return. I want to clock the NB as high as I can when still getting results, but if it's net result is simply increasing my temps I would rather keep it modest.

If my theory is correct the efficiency peak would be 2400MHz NB with DDR2-1200, although there would still be benefit of faster NB or faster memory. (more bandwidth -> lower latency)

As you see from Tony's table, pi calculation improves along with NB despite memory lagging. So if what matters is the traffic between core and L3, the higher NB the better the performance.

On the column of tables, we see the performance flattens with DDR3-1333 past NB frequency 2.8GHz, which is rectified by DDR3-1600. So it tells me DDR3-1333 is not feeding enough data to NB running @2.8GHz. I'd assume DDR3-1600 would be good up to NB 3.2GHz and at that point we would see diminishing returns once again..

[Enoc]

tony can you test cas 5 at 1333 ?

[Tony]

tony can you test cas 5 at 1333 ?

cas6 at 1600 is easier but the results will be the same...

higher the NB clock the more performance...2600/2800 is enough for 1333, 2800/3000 is enough for 1600.

Silly think is 2600NB is dead easy...why AMD don't preset the NB to 2600 i have no idea

[chew*]

tony can you test cas 5 at 1333 ?

I tested cas 5 1400+, its fast.......but does not scale as well....I think your better off with tony's reccomendations

[Tony]

Updated to include HT clocking and 3D

[roofsniper]

nice funny to see that HT overclocking lowers the scores. doesnt make sense to me but w/e. im gonna try underclocking my HT and see what happens. also btw the left side of the HT graph is messed up. should say score instead of HT overclock.

[Tony]

nice funny to see that HT overclocking lowers the scores. doesnt make sense to me but w/e. im gonna try underclocking my HT and see what happens. also btw the left side of the HT graph is messed up. should say score instead of HT overclock.

check again

[roofsniper]

check again

lol cool. got an explanation for why the scores get lower as the speed goes up? also you planning to do any more tests?

[Tony]

will try quad Xfire next to see how that does...if its the same Sampsa is wasting his time clocking HT high for his 3d runs

[roofsniper]

just ran the canyon test of 3d mark at 1400 1600 1800 and 2000 and all of those were normal on my 4870. scaled linearly from 1400 to 2000.

[angoholic]

very nice work

[Tony]

Tested 2x 2870X2, No gains with HT, tested PCIE OC single card to 125, tiny gain, with HT OC no gain. Dual cards PCIE 115 no gains, with HT no gains.

Weird thing is 115PCIE would not boot with HT 2000, at 115 dual cards but would run OK with HT at 2600...i think there is something wrong with bios on the Asus...at least PCIE and SATA are decoupled as the SSD had no issues even at 125 PCIE.

Not got good single cards...im wondering if 2x 4890's would gain thru PCIE OC and then HT OC on top.

[roofsniper]

well the ht thing is interesting that as you overclock it it loses performance. from the simple things i did from 1400-2000 it gained every step. i have seen others lose performance too by overclocking but maybe if i push it past 2000 i would start to see performance drops.

[Tony]

Lets go back a yr, 780i HT link between NB and SB, running 2600 or 2800HT along with PCIE at 141 with dual cards made a massive gain...BUT things were different.

1 the NB supplied 16X and the SB supplied 16x PCIE so an overclock to the PCIE bus would yield gains if the bus connecting the domains had a frequency increase.
2 790FX supplies both 16X slots, so the ONLY gain we should see with the video cards would be PCIE clocking until the transfer of data from PCIE to HT was the bottleneck then an increace in HT will be needed from the NB to the CPU.

So...why does increasing HT lower 3D performance...if anything an increase in HT bus should do nothing if the bus is already not the bottleneck?

Something is not right here

[roofsniper]

really i dont get how overclocking can ever lower performance. i know people have had issues with running the nb too high sometimes and getting less performance and just like you and others have tested with the ram. even if u run 1600 at c6 its still slower than 1333 c6. this phenom II platform has just been confusing.

[demonkevy666]

Lets go back a yr, 780i HT link between NB and SB, running 2600 or 2800HT along with PCIE at 141 with dual cards made a massive gain...BUT things were different.

1 the NB supplied 16X and the SB supplied 16x PCIE so an overclock to the PCIE bus would yield gains if the bus connecting the domains had a frequency increase.
2 790FX supplies both 16X slots, so the ONLY gain we should see with the video cards would be PCIE clocking until the transfer of data from PCIE to HT was the bottleneck then an increace in HT will be needed from the NB to the CPU.

So...why does increasing HT lower 3D performance...if anything an increase in HT bus should do nothing if the bus is already not the bottleneck?

Something is not right here

it's not just connected with nb/L3 cache between cores ???

there is some inter connecting Ht between cores right.

HTT speed increase on should change those, but if can't by pass the nb/l3 cache it isn't going to improve anything.

anyone got some old bios to turn off the L3 cache ? maybe then we can see lol

[imamage]

Looks like I can keep my NB speed high enough isn't bad idea

[roofsniper]

it's not just connected with nb/L3 cache between cores ???

there is some inter connecting Ht between cores right.

HTT speed increase on should change those, but if can't by pass the nb/l3 cache it isn't going to improve anything.

anyone got some old bios to turn off the L3 cache ? maybe then we can see lol

yea i think when you overclock HT it also overclocks the HT connecting the cores as well.

[chew*]

Thank you much again, Tony. It's exactly what I wanted to see. I think I'm starting to get to know about PII. And probably you've known about it all along. :P My hypothesis regarding PII architecture from your tables above:

1. When data can be replenished from L3, K10's performance keeps increasing to NB frequency, until it reaches the same frequency as the core frequency.
2. When data needs to be fetched from system memory, increasing NB frequency alone will give diminishing returns unless memory can keep up with it. Ideally it should be [NB frequency = 2 x memory frequency] (dual-channel)

So in my imagination, a 'perfect' 955BE would be 3.2GHz core / 3.2 GHz NB / dual-channel DDR2-1600 or higher.

Does this make sense to you, or am I totally wrong?

Your theory would be wrong.

I tested equal NB to cpu ratio.......the NB must not match or exceed cpu frequency or it takes a rather large performance hit.... for whatever reason it is not happy with a 1/1 ratio......give me a few and I will show you...

Once again I hope you don't mind Tony

For the sake of saving my poor CPU I used at PH II 925 for example.....

So here we have 1/1 ratio with CPU and CPU/NB efficiency = 67656 ( result time multipled by cpu speed ) <--lower is better



And here we drop the NB down to 2600....This should be slower in theory..... efficiency = 67436 ( result time multipled by cpu speed ) <--lower is better



And here is where the Real sweetspot truly is...... efficiency = 67411 ( result time multipled by cpu speed ) <--lower is better

[eRazorzEDGE]

i wonder chew*, does it continue that trend as you raise the CPU frequency? not sure how high you can clock your NB

[chew*]

i wonder chew*, does it continue that trend as you raise the CPU frequency? not sure how high you can clock your NB

I would be glad to do it on phase but.....thats not going to help you guys on air /water....Trying to avoid that as although its fun....its not much of a help to those on air/water.

I managed to test it on my 940 at 3000+ at one point, I know with ddr II it took an even bigger performance hit.

[massman]

From my testings, I found that the higher the NB clock frequency, the more performance up untill your NB is clocked higher than the cpu itself. At that point, the performance takes a real hit.

Thanks for the insights regarding the memory overclocking. It seems that we'd all be better off using low latencies rather than high frequencies, which is pretty much the same I noticed when comparing the AM2 platform to the AM3 platform. Highly overclocked DDR2 memory (620+) combined with tight timings (CL4-4-4) will be able to match any DDR3 configuration with timings equal to CL7-X-X and looser. This reminds me of the Msi P45 DDR3, which overclocked better at DDR2 timings!

Regarding the HT Link findings: are you sure nothing went wrong during the tests? I've done the EXACT same tests in combination with my 4870X2 card and an increase in HT Link frequency showed me a clear increase in 3D performance (3DM01 - Nature). This HT Link frequency has, in addition, given me fantastic performance increases when benching the HD3300 IGP at 1.1GHz.

[Nedjo]

Thanks for the insights regarding the memory overclocking. It seems that we'd all be better off using low latencies rather than high frequencies, which is pretty much the same I noticed when comparing the AM2 platform to the AM3 platform.

well, bandwidth still do counts... games do benefit from BW

[massman]

Sure, but the bandwidth is just the same:

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

Check the Everest chart and compare GD70 to the M2RS in the 3.6G test setting. 600 CL4 beats 800 CL7