|19-11-2009, 09:15 PM||#1 (permalink)|
Join Date: Nov 18 2009
Article: CPU VS GPU
Some years ago i start a tuning project called Immolator. It was based on old Abit board VH6-T, Pentium III-S 1400 "tualatin". I was buying old parts so it was quite cheap. The goal was not to reach best FSB overclock but to archieve best CPU, graphic and memory performance measured mainly by Everest, 3dMark2001 and 2003.
CPU, Board and RAMs were excellent pieces of its time. Overall CPU and memory performance was more than 10 percent better than other scores available on ORB, even in comparison with famous Asus Tusl.
And now lets move to graphics...
My first graphic was Ati Radeon 9800 Pro. Modded piece with really high performance. The overall score in 3dMark 2001 was 11 000pts, in 2003 5600 pts. The main limit of this card was at FSB/AGP lock... At 160mhz FSB i was unable to raise fsb without crashing the graphics.
Later i bought radeon X850xt. This card had PCI-E/AGP bridge chip. Card was impossible to crash due high AGP frequency. Overall score in 3dmark 2001 was again 11 000pts, in 2003 average was 9500pts. My personal record is more than 11000 points.
At time when i bought this card people were arguing that graphic card is too strong for the machine. I felt that I am hitting some sort of limit and in games like Oblivion will performance be same, mostly because of memory.
For few months i had best score with Pentium III in 3dmark 2003. Then someone picked HD 3800 and similar board. 18000 points.My project was trying to hit the limit of CPU, not the graphics, and one the goals was to measure downlimit of CPU/GPU coexistence.
When i changed the graphic card for the first time the score on 3dmark 2001 did not changed. That shows the downlimit CPU/GPU coexistence. This 3dmark is more CPU based while graphic effects are for these days quite basical. Adding more GPU performance did not affected score and FPS.
At all i was unable to hit the upper limit represented by 3dmark 2003. Even hi-end card produced in year 2005 (x850) was unable to fill all graphic requirements of this benchmark and thus to process all CPU requests "in time".
But there is also one important value - effective usage of CPU/GPU. I performed both 3dmarks with Vertical Sync and triple buffering. The scores were significantly lower in both benchmarks and both cards.
This gived me an idea for a measurement process of CPU/GPU efectivness.
1. All depends on applications which you use.
2. You need two graphic card and two benchmarks which are measuring game performance. (i prefer those two 3d marks).
3. When you raise CPU performance, you are automatically giving more work to GPU (per second).
4. Every FPS higher than refresh frequency of your display is not displayed.
A. Hitting the Downlimit
Run the both cards with less intensive test (3dmark 2001). When both cards give same results on default settings you can use them for further measurement. If not remove the slower graphic card and try something faster and repeat.
B. Measuring the Uplimit
Test both cards with more intensive test (3dmark 2003). If both of them is giving same results change the one of them for something stronger. If there is nothing stronger on market you hit the limit at least at this time If even stronger card is giving same results you hit the uplimit with all cards used.
C. Measure the Efectivness
Test both cards with more intense test with Vertical sync and Triple buffering on.
D. Check the results
Now look on the numbers. Downlimit for both cards should be same. Uplimit should be better for stronger cards. Compare the uplimit results with Efective results.
If the results for uplimit and vsynced values are same CPU and GPU are both unable even hit FPS equal to refresh frequency of your display. One of the components is too old, and you must know which one is it. You can try to lower the refresh rate of you display to minimum and try again efectivness test.
If Vsynced rates are significantly lower than normal values it is certain that the limit is your display. Measurement was then sucessful. FPS does not matter if you cannot display it, but graphic card can use up this "wasted" performance for better AA, AF, HDR or other effects.
Since downlimit for both cards is same it is certain that no graphic card (strong or weaker) is affecting CPU - causing it to slow down and wait for GPU for basical graphic operations. The bigger percentual difference between Uplimit value and Efectivness value is, the more percent of GPU is available for improving image quality, and other GPU usage without negatively affecting game and CPU performance.
If you test two graphic cards on same machine and one of them is performing worse it means that CPU is negatively affected by it. In any case this graphic card will lower performance of your CPU. Do not use and replace it with stronger card.
If you test two graphic cards on same machine and both of them are performing same, even if it is certain that one of them is significantly stronger you do not need upgrade you graphic card any further. It will not improve the system performance any further. This limit is hard to hit. In this case you can safely use the weaker card.
CPU/GPU/Display efectivness (50-100Hz)
This value shows how much performance can your system display with any graphic card. Mostly the display (crt, LCD, other) is the limit. You can measure it by running test when you turn on the Vertical Sync and Triple buffering on your graphic card. Then make a percentual comparation with GPU/CPU uplimit.
Percentual difference between Uplimit value and Efectiveness value.
Uplimit = Efectiveness + >100 percent : Graphic card is too strong. Probably you would never experience full performance of your machine. It is good to improve CPU, or display (but only in case you own 20 years old compaq with 50Hz refresh).
Uplimit = Efectiveness + 100 percent : Half of graphic performance can be used for generating FPS, rest is free for any other use.
Uplimit = Efectiveness + 75 percent : Large portion of graphic is still available to system. You might turn all effects on without even be able to see FPS downgrade.
Uplimit = Efectiveness + 50 percent : Some portion of graphic is still not consumed by CPU and can be used for effects, but you may rarely experience framedrops due graphic card.
Uplimit = Efectiveness + 25 percent : Everything should work fine, but do not play with effects.
Uplimit = Efectiveness + < 25 percent : Graphic card too weak or display extremely strong. I would recommend upgrade of graphic card.
Ideal difference between Uplimit and Effectiveness is somewhere at 62 percent. When this value is higher all framedrops you will see are caused only by CPU, memory, HDD and other possibly slow processes. If efectiveness is lower you might be able to see framedrops caused by GPU when you run GPU based effects.
For Downlimit test use sofware depending on your CPU. If you have dualcore be sure that its supports multiprocessing. Use previous versions of DirectX ( 9 or older).
For Uplimit test use software depending more on your graphics. Use the most recent test supporting the latest directX. Also be sure that this test will support multiprocessing
Also please note that this test is based on model which counts with situation where you have Triple buffering and Vsync on at all times, while FPS lower than refresh frequency is considered as measurable, and FPS lower than 25 is considered as critical framedrop. If FPS = Refresh rate you are displaying 100 percent of content generated by your machine and any higher FPS cannot be displayed corectly, but GPU performance unused in FPS generation can be used elsewhere without lowering FPS.
Last edited by Offler; 19-11-2009 at 09:39 PM.
|article, cpu, gpu|
|Currently Active Users Viewing This Thread: 1 (0 members and 1 guests)|
|Similar Threads for: Article: CPU VS GPU|
|Thread||Thread Starter||Forum||Replies||Last Post|
|DFI Lanparty UT X58-T3EH8 LGA1366 Photos, Bios and Overclocking Information Guide||eva2000||DFI Intel Motherboard / CPU||350||01-10-2010 10:27 AM|
|hwbot i4memory.com team challenge - Jan - Dec, 2009||eva2000||HWBOT Official Rankings||1176||14-02-2010 12:43 AM|
|Initial notes, findings and tips DFI UT X58-T3EH8||eva2000||DFI Intel Motherboard / CPU||0||04-08-2009 04:17 PM|
|Gigabyte GA-EP45 Extreme rev1.0 Review||eva2000||Gigabyte Intel Motherboard / CPU||34||15-03-2009 08:53 AM|
All times are GMT +11. The time now is 07:17 PM.