Actual Tweak Results

[keithlm]

After getting tired of reading people post things about how certain things "don't matter" and a reviewer need not worry about tweaking or using the appropriate memory speed I decided to do my own tweak benchmarks. I did this mainly for my own edification, but thought I'd post the results for the benefit of anyone curious.

I only ran them one time each... so this isn't an end-all review. Nor do I profess to call myself a "professional reviewer". Of course to be completely thorough I should also do all of these same tests at the fastest CPU speed I can use. But I'll leave that to the professionals: My little effort revealed exactly what we would expect.

(STOCK below means 3.0Ghz Phenom II 940, un-ganged, 5-5-5 timings, stock video. Fusion for gaming only adjusts the services. It is not overclocking the CPU or the video card.)

I ran 3DMark Vantage with several changes:

STOCK system with DDR2-800
STOCK system with DDR2-1066
STOCK system with DDR2-1066 using ATI Fusion for Gaming
NB@2.4Ghz with DDR2-1066
NB@2.4Ghz with DDR2-1066 using ATI Fusion for Gaming

RESULTS:


It is easy to see that tweaking and using the appropriate memory speed in a benchmark could make a difference in the end results. Some might not consider the 4% or 5% that can be gained to be worth the effort. And remember: this doesn't include changing the CPU core frequency; this is only the gains that could be obtained by tweaking everthing else.

NOTE: I also did a run with ganged but the results were within 10 points... too close to actually call it a difference. (Margin of error and all.)

[demonkevy666]

you didn't try 4-4-4-12 800mhz with 2.4ghz NB ?

[keithlm]

you didn't try 4-4-4-12 800mhz with 2.4ghz NB ?

NB2.4Ghz
DDR2-800 4-4-4-12
(Using ATI Fusion to give it the best edge possible.)

P11478
GPU 11856
CPU 10476

(So the GPU is very close the the fastest. But the CPU speed is not as good. This shows that 4-4-4-12 DDR2-800 has roughly 2/3 of the gain of going from DDR2-800 at 5-5-5 to DDR2-1066 5-5-5.)

[Pointhore]

Thanks for taking the time to show us the difference. It's not a huge gain, but it is a gain worth mentioning.

[Titan7171]

Thanks for showing that these little changes can make a difference.

I always run my system tweaked already for 24/7........I have a score for this setup and then I run a extreme setup for my highest score.

[keithlm]

Realizing that 3dMark Vantage is not a very good result for this NB comparison, I repeated the test in Phoroniz-Test-Suite. (CPU at 3.6Ghz, memory at DDR2-1066 5-5-5-15 and the NB at 1.8Ghz and then at 2.4Ghz. Obviously I don't have ATI Fusion for Gaming on Linux.)

You can see the comparison at: http://global.phoronix-test-suite.co...8369-7544-1128

But here are the tests with results that are more relevant. The red results show where the test results were opposite of the desired results. The results reflected what I basically already knew. This does show you want to use the fastest NB possible but it is not worth making the system unstable to do it.

NOTE: I am not stupid enough to add the percentages and divide by the number of test shown. That doesn't reveal anything statistically. (For the people that don't understand that and actually think it means something the number is about 43%.)

[SEA]

Wow!
I never thought NB matters THAT much!..

[keithlm]

Wow!
I never thought NB matters THAT much!..

Well I did only show the results where there was a difference... there were a lot of benchmarks where it made no difference at all.

(Actually I'm happy because many of them are identical. Which shows that everything else is basically equal between the tests.)

[Lightman]

Thanks for taking time to ran all the tests.

Interesting results from Linux. This answers why AMD needed Shanghai so much!

[Calmatory]

Now AMD needs CPU's with 3.6 GHz core/3.6 GHz NB.

Find it very odd though, 1.8 -> 2.4 is only 33 % improvement, yet there are results which show far better than 100 % scaling, e.g. the first 3 ones.

[Meizuman]

Not exactly the right place for this... but I got an idea why AMD dropped the NB speed to 1,8GHz with Deneb... Maybe it has everything to do with the fact that they tried to make it more compatible with the low-end mobos out there. And there is no problem for enthusiast to raise it higher with better mobos...

[roofsniper]

Not exactly the right place for this... but I got an idea why AMD dropped the NB speed to 1,8GHz with Deneb... Maybe it has everything to do with the fact that they tried to make it more compatible with the low-end mobos out there. And there is no problem for enthusiast to raise it higher with better mobos...

the HT speed will run lower if it is in a lower speed board. thats why am3 chips which run at 2000mhz HT can work in am2 boards which don't even support HT 3.0. im not exactly sure why amd had the HT speed lower on the 940 and 920 one reason why could be because it is one more reason to get the faster am3 version. either that or they were having issues with it at the time somehow.

so how can you get lower scores with having a higher nb speed? i see in some cases you even get a 30% performance loss by having a faster nb speed. what is the reason behind this?

[akaBruno]

Is this where intel beats them out?

[keithlm]

so how can you get lower scores with having a higher nb speed? i see in some cases you even get a 30% performance loss by having a faster nb speed. what is the reason behind this?

I'm still looking... I'll re-run the test at 1.8Ghz and 2.4Ghz just to make sure the result is not anomalous. (I'm re-installing the test suite now... they just released a new update. Plus I just ran a 3.0Ghz CPU run so I could compare with my old 9850 at 3.0Ghz. And also to compare to the Q9650 3.0Ghz results available on Phoronix. AND also to compare the stock 3.0Ghz Deneb to the i7 results.)

NOTE: Generally Phoronix is good about running the test 3 or 5 times to make sure you get an average. But I want to actually re-run that single test after manually changing the hardware speeds.

[roofsniper]

yea im not doubting it it just seems weird that clocking something higher would give you lower performance.

[keithlm]

The system did some updates to the OS this morning after I ran all of the other benchmarks... including some library updates to the main core.

I rebooted using the same kernel as I had used before... and recompiled the applications and then re-ran the FIO test... fileserver load. The results are slightly different. (I would need to re-run everything due to the OS changes... but this was the only one with the major anomaly. Well this and the bonnie delete test. I'll re-do that one now and then add the results to this post in a few minutes. Again I'll need to recompile that test for the same reasons.

NB 1.8: 85.74
NB 2.4: 83.94

SO... the faster NB is slower... but only by 2%.

EDIT: Actually phoronix test suite removed the BONNIE test from the suite. I will not revert to the older version of the test suite just to re-run the test. (I guess they had other problems with that test.)

[demonkevy666]

agena L3 caceh 32 way 2mb,2000mhz
Deneb L3 cache 48 way 6mbs 1800mhz

deneb at stock speeds only shows 5% improvement even with it lower. increasing the speed probably won't make much of a difference unless the cores speeds up too to use all the nb speed and cache.

it was much easier for agena to use the nb speed since it's l3 cache was small for it.

[gOtVoltage]

The system did some updates to the OS this morning after I ran all of the other benchmarks... including some library updates to the main core.

I rebooted using the same kernel as I had used before... and recompiled the applications and then re-ran the FIO test... fileserver load. The results are slightly different. (I would need to re-run everything due to the OS changes... but this was the only one with the major anomaly. Well this and the bonnie delete test. I'll re-do that one now and then add the results to this post in a few minutes. Again I'll need to recompile that test for the same reasons.

NB 1.8: 85.74
NB 2.4: 83.94

SO... the faster NB is slower... but only by 2%.

EDIT: Actually phoronix test suite removed the BONNIE test from the suite. I will not revert to the older version of the test suite just to re-run the test. (I guess they had other problems with that test.)

Its tricky but sometimes the Higher NB needs simply more Volts to perform.

Sometimes if i try Running High NB without raising the NBvolts to where it needs to be (to keep it stable) ,i have seen the NB run slower in apps even though my SuPi is faster.

Thus i run 1.65volts for 2600+ Nb (9950)..Theres Of coarse another factor (Heat )will slow the NB down too . Usually the NB will slow when Raised to a point in which the Aplication overheats it (from load.) This means your running to much Voltage or are just above your optimal NB speed before it goes down hill. Ive noticed in Some Programs/apps that use the Ram more heavily can load the NB when its overclocked too much or not enuff volts have been used to keep stable,,

In all ,,find the NB speed that runs Everest with the Highest NB speed possiable. When or If i go just a hair Above this speed,, i will see the performance drop off (the NB will lose performance). Its usually because your past the optimal MHZ in most cases ,or it overheated. Video cards do the same thing when Overclocked too much actually.

Take the time to find a sweetspot,, I usually lower the NB just below the Max i can Overclock it . Usually i will not lose any performance in most apps with it Being @ 2600mhz for me.


Note:
My max with 9950 has been 2730mhz on the NB..Above 2700mhzNB actually would run slower than if i had it set @ 2670mhzNB. hope that all made sense.

Ganged gave me worse performance @ 5-5-5-12 1066mhz while using 2600nb vs Unganged when using High NB and 4-4-4 -6 @ 1066mhz.

All these little things help the NB speed.. matter of fact Ganged usually resulted in having to run looser timing and less NB overall once i went above 3.5ghz.

Again this is on 9950(940 should be no differnt except for HT/Nb default speed 1800 vs 2000) 940 can run 2800 -3000mhz NB. Once its tweaked right you should see a much better performance increase in Most aaaapps when raising the NB Trust Me

I cant wait for 945 to come around.. torture waitng.

DEMON has a good point on the Cache above this post You should still be able to increase overall performance even more when the CPU goes higher. Just tweak that thing . Having NB above 2600 on C2 will show better results. Moe Cach means you need to speed up the NB to utilize it. This even means tighning or loosing the ram in Ganged or Ungaged to help the NB run smooth as well.

My max results are with the HTT/HT and NB at the Same speed. This has given me the best possiable performance ive ever had on B3 hands down.

EXAMPLES BELOW,,

L@@K
Ram is @ 5-5-5 15


L@@K
Ram is @ 4-4-4-15

[roofsniper]

nice post and well said it makes sense now.

[gOtVoltage]

Cant wait to get me hands on a 940 or 945...It diffacult to hold out sometimes. It payed off for my 9950,, i passed on the 9850...even though i knew both would be a good choice atm.

Now im doing the same with the 945 i think. So Febuary is only a month now roughly. I really want to see what a AM3 part will do in this Mobo. I planned this Mobo for 940 but the 945 seems to be on par for Feb .

[keithlm]

I did the universe-x test at Phoronix-test-suite. (I also updated to the newest Linux and the newest Phoronix-test-suite.)

EDIT: Again with 3.6Ghz on the CPU and the NB at 1.8Ghz and then 2.4Ghz
http://global.phoronix-test-suite.co...14-15446-24502

The universe-x test includes a bunch of games and other graphical programs.

Most of the results are again about the same... many show gains. One showed worse results. Go figure.

BTW: I can't increase the NB VID on this motherboard. The bios has what is called "NB Voltage Control" which defaults to 1.815V. In WINDOWS I can adjust the NB VID with AOD but not in Linux.

NOTE: For people that don't realize it... when you load Phoronix-test-suite... it loads all of these games. These are the actual games not just synthetic benchmarks. So in addition to getting a benchmark platform you also get a bunch of games. Some are not too bad either.

[roofsniper]

im guessing that test was run with the nb at 1.8 and 2.4 like the other tests?

this nb speed thing is kinda odd tho. gotvoltage explained it pretty well but its still weird that higher nb can be worse. is there anyway of finding the sweetspot or do u just have to tweak around till you get it?

[keithlm]

im guessing that test was run with the nb at 1.8 and 2.4 like the other tests?

this nb speed thing is kinda odd tho. gotvoltage explained it pretty well but its still weird that higher nb can be worse. is there anyway of finding the sweetspot or do u just have to tweak around till you get it?

Sorry... but you are correct it was with the 1.8 and the 2.4ghz NB. (3.6Ghz CPU) I edited the above to show that.

Actually I just ran a full universe and the same test (openarena) is better at CPU 3.0ghz/NB1.8NB than it is at CPU3.6Ghz/NB2.4NB so it there is definitely something with that one test that doesn't like the NB speed. The rest of them are fine. (I ran out of disk space so it failed a bunch of the tests so I won't post the results yet. I had also played with using ext2 instead of ext3 so I don't trust the results. (Besided changing more than one variable at a time == bad science.)

I wish this bios would let me adjust the NB voltage.

[roofsniper]

you think it might end up being better with newer bios updates? since phenom II is pretty much on beta bios for everything.

[keithlm]

you think it might end up being better with newer bios updates? since phenom II is pretty much on beta bios for everything.

If a newer bios would allow me to change the NB and HT speed... and increase the NB VID... then it would be MUCH better.

Right now I can change the NB speed. The HT will stay at 1.8Ghz and the NB VID is default.