More P5B secrets uncovered

[>trinity<]

how many of us work 24/7 out of windows ??

i guess that tests should be done where we are actually going to seek for maximum performance.

but since any of you doesn't trust everest, i'll post exacly the same settings as i posted with everest, and then we'll see if something changes.



As you can see, different programs equals to different results.

so i keep saying, test your own hardware and take your own conclusions.

i doubt that you have to reach 460+ to equal mem bandwith that you achieve with 399/400 whatever.

What happen to the pics? There is a pic missing from your earlier post as well.

Edit: pics are showing now.

[aggybong]

Great post, I never thought of using Mbench for the access latency but I will now

See if anything happens from 467 up also, thats where the next strap in theory should kick in.

I just redid the post with a bit more information. I could barely test 458 with my Rev2s at 4-4-4-12, so I don't think 467 would be possible with the same timings. I could try 5-5-5-15 at 466, 467, and 468, I suppose, but at those speeds my CPU might cause troubles. Ah well, off to try anyway!

[Tony]

I just redid the post with a bit more information. I could barely test 458 with my Rev2s at 4-4-4-12, so I don't think 467 would be possible with the same timings. I could try 5-5-5-15 at 466, 467, and 468, I suppose, but at those speeds my CPU might cause troubles. Ah well, off to try anyway!

All you need is cas5, the dimms do a lot better cas5 than cas4, you may also want to test 5-4-3 instead of 4-4-4, less voltage and the same performance ot better from my tests.

[aggybong]

What does Static Read Control do?

[Tony]

What does Static Read Control do?

Not sure yet, hard to see what register its setting. Maybe someone else can chirp in with some answers.

[ScottFern]

So I am having trouble interpreting these results. Is the conclusion its better to stay closer to 400-450FSB and use the default multiplier than say 500X7, where 8 is the default multi?

[aggybong]

I'm not seeing any difference in the 466-469 range.

Memtest86 Results @ 5-5-5-15-5-42:
7x466: 4854
7x467: 4863
7x468: 4875
7x469: 4886

Sandra 2007 @ same timings:
7x466: 7539 / 7565
7x469: 7562 / 7563

I wanted to try 8x46*, but my CPU wouldn't have it at 1.45v, so I said forget it Do you have any advice on the TRRD, Write to Read Delay, Read to Precharge Delay, and Write to Precharge delay? At the moment I just leave them at 10, but that's probably not optimal.

I'm done testing for now, though! Thanks for the advice, Tony.

[>trinity<]

I just did 3 bandwidth tests with sandra 2007.

7879 = 400fsb bios
7889 = 403fsb from 400-403 used clockgen
7375 = 404fsb bios

Here are the tests.

http://www.trinity.hdc.net.au/1118.jpg
http://www.trinity.hdc.net.au/1119.jpg
http://www.trinity.hdc.net.au/1120.jpg

[bachus_anonym]

I think, this is another good way to show what Tony and FCG are talking about... To show that past 400MHz there is a "dip" in performance, I decided to complete few SuperPi 32M runs, a great bench that can uncover fluctuations in bandwidth and latency.

First, I booted at 400MHz and run 32M at three ratios - 533, 667 and 800. Then booted at 402MHz and did the same. At the end, to show that "dip" may be somewhat reversed (up to the point, unfortunately), I booted at 400Mhz and then once in Windows, raised clock to 402Mhz with ClockGen.

Boot @ 400x9        Boot @ 402x9        Boot @ 400 & CLG to 402

DDR2-800  --> 14:10.969  DDR2-804  --> 14:42.922  DDR2-804  --> 14:06.672
DDR2-1000 --> 13:35.765  DDR2-1004 --> 14:07.297  DDR2-1004 --> 13:32.562
DDR2-1200 --> 13:17.109  DDR2-1206 --> 13:51.891  DDR2-1206 --> 13:13.109

Interesting, eh? Unfortunately, raising clock past ~430Mhz using ClockGen, ends with a crash


EDIT: Oh, yeah... BTW, this is what happens when one compares 516x7 and 400x9, with keeping memory speed as close in both cases as possible:

516x7=3611MHz, DDR2-1032 --> 13:55.625   400x9=3600MHz, DDR2-1000 --> 13:35.765

[Vapor]

I think, this is another good way to show what Tony and FCG are talking about... To show that past 400MHz there is a "dip" in performance, I decided to complete few SuperPi 32M runs, a great bench that can uncover fluctuations in bandwidth and latency.

First, I booted at 400MHz and run 32M at three ratios - 533, 667 and 800. Then booted at 402MHz and did the same. At the end, to show that "dip" may be somewhat reversed (up to the point, unfortunately), I booted at 400Mhz and then once in Windows, raised clock to 402Mhz with ClockGen.

Boot @ 400x9        Boot @ 402x9        Boot @ 400 & CLG to 402

DDR2-800  --> 14:10.969  DDR2-804  --> 14:42.922  DDR2-804  --> 14:06.672
DDR2-1000 --> 13:35.765  DDR2-1004 --> 14:07.297  DDR2-1004 --> 13:32.562
DDR2-1200 --> 13:17.109  DDR2-1206 --> 13:51.891  DDR2-1206 --> 13:13.109

Interesting, eh? Unfortunately, raising clock past ~430Mhz using ClockGen, ends with a crash

Holy cow that is drastic....very good info though, thanks Bachus

[dinos22]

let's see other 965 tests with SPI

i have a feeling it's more drastic on P5B Deluxe but could be wrong.....just thinking about some SPI results i was getting DS3 vs P5BD hmmmmmmmmm

[aggybong]

Go forth and find willing DS3\DQ6 users :P

You know, I find it odd that when comparing 7x458 and 8x400 using MBench, that the 7x458 configuration loses everything BUT "Write Datarate (SSE)", which it wins by over 1000 Mb/s. Isn't that rather large?

[Tony]

let's see other 965 tests with SPI

i have a feeling it's more drastic on P5B Deluxe but could be wrong.....just thinking about some SPI results i was getting DS3 vs P5BD hmmmmmmmmm

I have a DS3, will test next week to see what they are upto also

Remember asus=gigabyte now, so they may have the same tweaks

[eva2000]

it all depends if Everest actually measures anything, or calculates bandwidth by bus speeds etc.

memtest measures it as far as i know

does memtest86+ v1.65 work on 965p to show bandwidth ? as on my badaxe and p5wdh memtest reports my bandwidth anywhere between 28,000MB/s and 68,000MB/s LOL

[dinos22]

you know with this whole asus=gigabyte thing i reckon the rivalry between these departments is probably even greater than before heheheh doubt they'll share all the info among eachother knowing how other big companies work (at least in australia hehehehe)

[dinos22]

does memtest86+ v1.65 work on 965p to show bandwidth ? as on my badaxe and p5wdh memtest reports my bandwidth anywhere between 28,000MB/s and 68,000MB/s LOL

yeah it does

[aggybong]

I'm not even using Memtest86+, I'm using this: http://www.softwarecove.com/memtest/

[Tony]

bachus

Here is the test for you. Clockgen up to 427 from 400 fsb and run 32M pi, then boot in at 427 and run 32M pi.

I bet we see a 1min faster run using clockgen

Someone needs to do this under LN2, i bet with the right tweaks we could see 8sec 1M easy

[fscussel]

I think, this is another good way to show what Tony and FCG are talking about... To show that past 400MHz there is a "dip" in performance, I decided to complete few SuperPi 32M runs, a great bench that can uncover fluctuations in bandwidth and latency.

First, I booted at 400MHz and run 32M at three ratios - 533, 667 and 800. Then booted at 402MHz and did the same. At the end, to show that "dip" may be somewhat reversed (up to the point, unfortunately), I booted at 400Mhz and then once in Windows, raised clock to 402Mhz with ClockGen.

Boot @ 400x9        Boot @ 402x9        Boot @ 400 & CLG to 402

DDR2-800  --> 14:10.969  DDR2-804  --> 14:42.922  DDR2-804  --> 14:06.672
DDR2-1000 --> 13:35.765  DDR2-1004 --> 14:07.297  DDR2-1004 --> 13:32.562
DDR2-1200 --> 13:17.109  DDR2-1206 --> 13:51.891  DDR2-1206 --> 13:13.109

Interesting, eh? Unfortunately, raising clock past ~430Mhz using ClockGen, ends with a crash


EDIT: Oh, yeah... BTW, this is what happens when one compares 516x7 and 400x9, with keeping memory speed as close in both cases as possible:

516x7=3611MHz, DDR2-1032 --> 13:55.625   400x9=3600MHz, DDR2-1000 --> 13:35.765

ok, I have came to a conclusion of my own... when you use clockgen or another "windows" program to change the clock, apparently you cheat the board in a way that it doesn't apply the next strap, so you keep the best of both worlds, a faster strap and an fsb/clock increase... that has lead me to another point: probably when rising the FSB with clockgen you will reach a crash/limit must earlier than by doing that using the bios... so, let's say, if you can reach 500mhz by setting it in the bios, you will only reach 430mhz when using clockgen, that's because of the faster strap used when rising from windows (clockgen)... someone with hardware could test that !! I don't have my C2D rig yet...

[dinos22]

Someone needs to do this under LN2, i bet with the right tweaks we could see 8sec 1M easy

i don't know about that

you still need a CPU that can do 5.4Ghz

are you saying that 965 chipset prior to speedkink of >400+FSB would be faster than 975X motherboards

[ScottFern]

I have ran my own ghetto benchmarks. I have a E6400 Allendale CPU.

Sandra 2005 Memory Benchmark - My memory is 1:1

500X7@4-4-4-10 = 7278
7230

400X8@4-4-4-10 = 7303
7236

400X8@3-4-3-10 = 7387
7321

[dinos22]

ok, I have came to a conclusion of my own... when you use clockgen or another "windows" program to change the clock, apparently you cheat the board in a way that it doesn't apply the next strap, so you keep the best of both worlds, a faster strap and an fsb/clock increase... that has lead me to another point: probably when rising the FSB with clockgen you will reach a crash/limit must earlier than by doing that using the bios... so, let's say, if you can reach 500mhz by setting it in the bios, you will only reach 430mhz when using clockgen, that's because of the faster strap used when rising from windows (clockgen)... someone with hardware could test that !! I don't have my C2D rig yet...

that's exactly what it is.......it's self explanatory by the info provided

others have done this and they crash even earlier than 430

[bachus_anonym]

bachus

Here is the test for you. Clockgen up to 427 from 400 fsb and run 32M pi, then boot in at 427 and run 32M pi.

I bet we see a 1min faster run using clockgen

I'm on it and will be back in a few

UPDATE1: I was a bit too optimistic and dinos22 is spot on... I can't get my system to run 32M at 427MHz booted from 400MHz It shuts down the moment I start SPi calcualtion! I will try 415MHz and perhaps this will at least allow me go thru few passes...

UPDATE2: Unfortunately, 415Mhz was no-go either, therefore 410Mhz will have to do, Tony

Boot @ 410x8               Boot @ 400 & CLG to 410x8

DDR2-1026  --> 15:22.390          DDR2-1026  --> 14:39.859

That'd be 43s difference, enough to cry over

[dr_sharp]

This is fantastic. Thank you all for the results... I will contribute if I get a chance. Keep it coming.

[dinos22]

this whole debate over what CPU to get it pointless.........24/7 people will hardly notice these changes

benchmarking is a different story and as things are atm good E6600 with 975X mobo for air/water is the best way to go.............or P5BD+E6600+phase for the budget conscious

that's one PITA about Intel..........you HAVE to buy the Extreme chips to compete..........AMD NF4 was not as motherboard limited as these..........remembers how easy it was for the Ultra-D/Expert to boot and benchmark at 420+MHz HTT >> too bad about the cold bug that's all

maybe future revisions of motherboards will be able to perform better we'll see