stargate125645

Introduction:
In my attempts to make the top 30 computers here on overclock.net, I ran many benchmarks, making the smallest tweaks between each run in hopes of pushing my system as far as possible. I did eventually reach the 100k 3DMarks level and the top 20 computers when my score was submitted, but afterwards I was left wondering why some of my system tweaks affected each benchmark differently, and specifically by how much each benchmark score changed. I started out by just recording my benchmark scores for the various overclocks I had, but I realized that the data spread wasn’t good enough to make any concrete determinations, so I decided to go about this in a scientific method.

I set up my system such that all overclocks were an incremental percentage increase over the default speed, allowing any results to be more easily extrapolated to other systems. I used three CPU clock speeds (2.66 GHz, 3.00 GHz, and 3.34 GHz) and three GPU clock speeds (the memory and core were set to the same values of 500 MHz, 600 MHz, and 700 MHz). For each CPU speed I tested each GPU configuration, and vice versa, for the five main graphics benchmarking programs: 3DMark01 SE, 3DMark03, 3DMark05, 3DMark06, and AquaMark 3. This culminated in a total of 45 benchmark runs in all, the results to which are summarized in the following sections.

Test System:
ASUS P5W DH Deluxe Motherboard
- BIOS version 1707
Intel Core 2 Duo E6700 Processor
- All overclocking done in the BIOS
- Cooled with a Tuniq Tower 120 and Silverstone FM121
- Multiplier of 10
Corsair XMS2 DDR2 800 Model TWIN2X2048-6400
- Timings of 5-5-5-12 (same as rated timings at 800 MHz) for all speeds
- Divider of 1:1
- RAM voltage set to "Auto" in the BIOS
SoundBlaster X-Fi XtremeGamer Professional Fatal1ty Soundcard
- Driver version 2.9.7
- Creative Console Launcher set to "Game Mode"
Sapphire Radeon X1900XTX Graphics Card
- ATI Catalyst Control Center 6.11
- In the 3D Menu, Standard Settings in ATI CCC 6.11 are first set to "Performance," then Mipmap Detail Level is changed to "High Performance" (which changes the Standard Settings to "Custom Selection")
- All overclocking was done with ATI Tool 0.26

Results:
The stock speed for my test processor (Core 2 Duo E6700) is 2.66 GHz, so each CPU overclock was a true overclock. However, in order to achieve significant overclocks of my graphics card, I had to assume that the default speeds were 500 MHz/500 MHz for the core and memory. This allowed me to use a constant overclock percentage for both the core and memory for each scenario, again allowing the results to be more meaningful when extrapolated to other systems. In both cases, I did not overclock the components near their maximum, so stability and thermal issues were of no concern during these tests. All of the values and settings in the “Test System” section of this FAQ were kept constant so as to make sure the component overclocks were the only effects of the benchmark score changes.

The tables showing the raw data were too large to put directly into this FAQ, so I have attached a zip file of the two images, which can be found at the bottom of this post. The first table represents the changes in the benchmark scores with a constant GPU speed and a varied CPU speed, and the second the opposite. Each CPU overclock is approximately 12.5% more than the previous overclock, and each GPU overclock is approximately 20% more than the previous. (Please note that when I refer to CPU overclocks of 13% and 26% in the following analysis, I am actually referring to overclocks of 12.43% and 25.16%, respectively. The plots worked out better with round numbers along the x-axis, so I used those numbers when analyzing the data so that it would be easier for the reader to refer to back to the plots.)

Although the tables give a good idea of how the scores changed with each overclock, it is hard to determine specific trends from them, so I have also created two plots for each benchmark. The first plot for each benchmark shows the effects of CPU overclocking with a constant GPU core and memory clock, while the second shows the GPU overclocking effects with a constant CPU clock speed. I will discuss what these trends indicate as we go along.




For 3DMark01 SE, it is apparent that CPU overclocks have a larger effect than GPU overclocks on the overall score, as indicated by the fact that a 26% overclock in the CPU yields similar results as a 40% overclock in the GPU. Moreover, overclocking the CPU gives a proportional increase in benchmark score (i.e. the 26% CPU overclock gave about twice the score change than did the 13% CPU overclock), whereas the GPU overclock is not linear: The effects of GPU overclocking on the benchmark scores decrease as the GPU is overclocked further. That is to say, overclocking your GPU from stock to 20% will give more of an improvement than overclocking from 20% to 40%. This implies that there is a point at which further GPU performance increases will reach a point of diminishing returns.




3DMark03 shows a definite change in the way your overall score is computed. Both the CPU and GPU show very linear trends in their effect on the benchmark scores, with twice the overclock yielding roughly twice the benchmark score increase. 3DMark03 is also much more dependent upon the GPU, as a 20% increase in the GPU overclock resulted in 3 times the score increase than a 26% increase in CPU speed did.




3DMark05 indicates the same importance of the GPU over the CPU, with a 20% GPU overclock again providing nearly 3 times the score increase of a 26% CPU overclock. For 3DMark05, though, an interesting trend develops with the highest GPU overclock (700 MHz core and memory), wherein the subsequent increase in CPU speed results in nearly twice the score increase than did the two lower GPU speeds, however this effect appears to die down as the CPU speed increases further, much the same way that the GPU effect diminished with 3DMark01 SE.




3DMark06 had by far the most linear trends, with each set of benchmark scores lying on nearly identical paths. Like 3DMark03 and 3DMark05, the GPU overclock is much more important, again with the 20% GPU overclock providing roughly 3 times the score increase that the 26% CPU overclock provided. This implies that both 3DMark05 and 3DMark06 have very similar formulae used in calculating the total benchmark score.




Finally we come to AquaMark 3, where the overclocking results are very similar to 3DMark01 SE, in that the CPU overclock is more important, and that the GPU overclock effect diminishes slightly as the overclock percentage is increased. A 26% CPU overclock provides nearly the same benchmark increase as the 40% GPU overclock did.

Conclusion:
Put succinctly, 3DMark03, 3DMark05, and 3DMark06 all depend much more heavily on your GPU overclock than your CPU overclock, meaning that having a good GPU is much more important than having a good CPU. 3DMark01 SE and AquaMark 3 both exhibit the opposite trend, with the CPU being a more important component than the GPU.

It is also interesting to note that AquaMark 3 and 3DMark01 SE are the only two benchmark programs of the five tested that don’t exhibit linearity in the majority (if not all) of their trends, indicating that the GPU and CPU abilities are combined to form a product (multiplication) in the two programs’ equations used to determine your overall score.

In addition, with 3DMark01 SE having the highest 3DMark total of the four 3DMark programs, these results would suggest that overclocking your CPU would give you the biggest gain in score per overclock percentage if you are totaling up all of your 3DMarks, such as is necessary to make the top 30 computers on overclock.net.

I hope this FAQ has helped you understand the finer points of the major graphics benchmark programs, and how overclocking components, specifically the CPU and GPU, will affect each score. At the very least, I hope this FAQ will help you get that last ounce of power out of your system and push you into the top 30 computers!

Notes:
I understand that these results cannot be used to predict exactly the effect that GPU and CPU overclocks will have with different setups; however the general trends should still hold merit, regardless of the system, and that was the point of my tests and my writing this FAQ. I performed the tests as generically as possible so that someone could use these results to help tweak their system for better benchmark results.

I also realize that this FAQ leaves the question of the effect of memory and core speeds on benchmark scores unanswered. Unfortunately, such a test is nearly impossible to do because this relationship varies a great deal more depending upon the architecture of each GPU. Some GPUs are bottlenecked by their memory speed, and others by their core speed. In order to prevent this potential problem from affecting the results of the tests I conducted, I set it up so that the core and memory were both overclocked the same amount for each scenario, as I explained earlier. This allowed me to use a general GPU overclock percentage, rather than having to compare core and memory overclock separately for my GPU, which in turn would make my results inapplicable to other systems.

For the record, according to tests I did run with AquaMark 3, my increasing my X1900XTX core clock by 20% yields approximately four times the score improvement that increasing the memory clock by 20% gives.

Lastly, I encourage anyone else to run these same tests with their system for me to add to this FAQ. Credit will of course be given to you in full; all I ask is that the same scientific method and attention to detail be given that I gave. (I do, however, reserve the right to refuse to post other data if I do not think it is up to par.)

Attached Files

raw_data.zip (58.9 KB, 41 views)

__________________
Links:
[ Project Nighthawk (NEW!) | Project Ablative Armor | G15 LCD Program Archive ]
[ Intel Core i7 Voltage Limits | Overclocking Effects on Benchmark Scores | Explanation of LCD Terminology ]

stargate125645

I am supposed to link the photos back to their host site, and I will do that soon, but right now I have to get ready for a dinner date. Enjoy, all!

__________________
Links:
[ Project Nighthawk (NEW!) | Project Ablative Armor | G15 LCD Program Archive ]
[ Intel Core i7 Voltage Limits | Overclocking Effects on Benchmark Scores | Explanation of LCD Terminology ]

kevg73

wow, that must have been a ton of work... very very comprehensive thanks for all the advice on getting the best benchmark scores...

again, wow!!!!

Ihatethedukes

Going on percent change is a little misleading. You should really at the very least plot them as points change as well as what you have since it's a better indicator than percentage increase. The difference between 20% of a score of 5 and 4 is still only one point. But the 20% difference between 5000 and 4000 is 1000 points! 1000 points is very very significant in real world terms, 1 point isn't. Yes, my examples are extreme but hold true throughout.

__________________

(Dr.) Bottlenecks: Or, How I Learned To Stop Worrying And Just Buy the GPU
Bottlenecks Re-examined: Lightmark an OpenGL Benchmark
Importance of CPU and GPU in HL2: Lost Coast benchmarks.

stargate125645

I see your point, but I wouldn't call it misleading. I do have the raw data attached for anyone who does want to know the scores. I plotted them in terms of percentages so that way it would be easier to compare between benchmark programs. Percentages can translate but points cannot in that regard.

Someone also suggested doing the score stability, which is a good idea. I may pursue that if I get the time (this FAQ was done a while ago and I just now got the time to post it).

__________________
Links:
[ Project Nighthawk (NEW!) | Project Ablative Armor | G15 LCD Program Archive ]
[ Intel Core i7 Voltage Limits | Overclocking Effects on Benchmark Scores | Explanation of LCD Terminology ]

alawadhi3000

nice, thanks for the hardwork, REP+

__________________
GeForce 8800GTS 320MB/9600GT/8800GT

3DMark06 11518/10963/13622
3DMark05 17948/17446/23147
3DMark03 36902/36240/42768
3DMark01 49944/43675/53827

Total 116312/109324/133364

stargate125645

The images have now all been linked back to the host's photo page, as per their TOS.

__________________
Links:
[ Project Nighthawk (NEW!) | Project Ablative Armor | G15 LCD Program Archive ]
[ Intel Core i7 Voltage Limits | Overclocking Effects on Benchmark Scores | Explanation of LCD Terminology ]

Massive

very very nice review, I should give you a lot of credit for those very usefull information.

I'm getting this page bookmarked

stargate125645

Next step: Your homepage!

__________________
Links:
[ Project Nighthawk (NEW!) | Project Ablative Armor | G15 LCD Program Archive ]
[ Intel Core i7 Voltage Limits | Overclocking Effects on Benchmark Scores | Explanation of LCD Terminology ]

stargate125645

I think it's uncanny how linear the relationship for 3DMark06 is. I wonder if they will make a 3DMark07 and how it will fare. I will have to test it as well when it comes out!

__________________
Links:
[ Project Nighthawk (NEW!) | Project Ablative Armor | G15 LCD Program Archive ]
[ Intel Core i7 Voltage Limits | Overclocking Effects on Benchmark Scores | Explanation of LCD Terminology ]