As long as www.overclockers.com refused posting this topic, and some poeple asked for a sekond part, I posted it here 
In the first part, which is posted here: http://www.overclockers.com/articles1118/
I tried to explain that strange effect. Now with the help of Grey Williams. We together thought out this possible explanation. I could post it a month ago, but he said that he has sent this to someone in ATI to verify, but I didn’t get an answer neither from him, nor from ATI.
This is really hard to explain and even harder to understand. Even in my language (Lithuanian), so don’t complain about this.
Everything begins from that the video card has some sort of engine that detects miscalculated data and sends it back to recalculate. Why there is such a thing? It allows video card to work under worst circumstances... Instable voltage, high temperatures etc. And this allows the card to be cheaper, you can use cheaper components because it doesn’t have to work 100% correct.
This engine can work properly sending mistakes back till up to 5% is miscalculated. It’s also possible to calculate this number, but I didn’t bother to do that. Although I know that it’s somewhere around 5%.
What happens if there is over 5% mistakes? Try to overclock too much and you will start to see holes in walls (or something like this), this is the data that couldn’t be recalculated but was purged away.
Now lets go further, if there is 5% of mistakes, this means that 5% of data is calculated twice (1 time incorrect, and 2nd time to correct it). This means that some of GPU speed is used just for nothing. That’s why good PSU and cooling is important to get those 5% back!
And this is the hardest part... Why this effects minimal FPS more, then average still increases…
Now don’t take these numbers serious, that’s just to understand everything easier. Lets say that we oveclocked GPU, we don’t see any holes but we have 5% miscalculation or every 1/5 of the second. Average FPS is 100, and minimal is 20.
Then, the possibility to get an error then calculating the frame in 1/100 of the second (100FPS) is many time smaller than then calculation takes 1/20 of the second. The bigger the FPS is, the less it effects. Most of the frames on 100FPS are correct, only a small amount of them are recalculated, and recalculation only takes 1/100 of the second. Then with 20FPS, there is 5 mistakes per second, the 1/4 of data needs to be recalculated! And that recalculation takes 1/20 of s. And no one can say that error can’t appear then recalculating…
This explains the strange effect with extremely low FPS like 1-3. At this point there should be 1-3 FPS (not overclocked), after oveclocking, game simply freezes at this point. It takes 1-1/3 of the second to calculate it, and it gets error every 1/5 of second, this all goes into that this frame can never be calculated right = 0FPS! Sometimes this even goes into game crash.
Now lets go to something that is for sure, I can tell that the whole frame isn’t recalculated. Then you OC too much you see a hole in the wall this the part that had to be recalculated. All the objects are divided into such areas, and all this area is usually recalculated if needed. The one thing you should know, is that polygon count and FPS don’t effect number of areas. All this goes again, the more data there is, the more data in one area is, the bigger the possibility to get a miscalculation.
That’s all for now, I don’t have any more info and I hope that someone understood what I wanted to say.
Now lets go to letters, I received a bunch of them, and the most common question was how I measured it. So, 1st I used 3DMark2001 SE all game tests. How? Final 3DMark score is the sum of average FPS in all game tests. So I have the average FPS. Now lets go to min. I start “Fraps” and simply look for the lowest number. I also checked it in 3DMark03. And the 3rd is Sam 2, I simply looked for the lowest point, I didn’t have the average here. The easiest way is simply to play a game and look if the it works smooth.
All the tests gave same results, this shows that all this isn’t a mistake. You can write to me if you have anything to say. I try to reply to every email:
[email protected]
And I repeat once again: don't mind the numbers!
In the first part, which is posted here: http://www.overclockers.com/articles1118/
I tried to explain that strange effect. Now with the help of Grey Williams. We together thought out this possible explanation. I could post it a month ago, but he said that he has sent this to someone in ATI to verify, but I didn’t get an answer neither from him, nor from ATI.
This is really hard to explain and even harder to understand. Even in my language (Lithuanian), so don’t complain about this.
Everything begins from that the video card has some sort of engine that detects miscalculated data and sends it back to recalculate. Why there is such a thing? It allows video card to work under worst circumstances... Instable voltage, high temperatures etc. And this allows the card to be cheaper, you can use cheaper components because it doesn’t have to work 100% correct.
This engine can work properly sending mistakes back till up to 5% is miscalculated. It’s also possible to calculate this number, but I didn’t bother to do that. Although I know that it’s somewhere around 5%.
What happens if there is over 5% mistakes? Try to overclock too much and you will start to see holes in walls (or something like this), this is the data that couldn’t be recalculated but was purged away.
Now lets go further, if there is 5% of mistakes, this means that 5% of data is calculated twice (1 time incorrect, and 2nd time to correct it). This means that some of GPU speed is used just for nothing. That’s why good PSU and cooling is important to get those 5% back!
And this is the hardest part... Why this effects minimal FPS more, then average still increases…
Now don’t take these numbers serious, that’s just to understand everything easier. Lets say that we oveclocked GPU, we don’t see any holes but we have 5% miscalculation or every 1/5 of the second. Average FPS is 100, and minimal is 20.
Then, the possibility to get an error then calculating the frame in 1/100 of the second (100FPS) is many time smaller than then calculation takes 1/20 of the second. The bigger the FPS is, the less it effects. Most of the frames on 100FPS are correct, only a small amount of them are recalculated, and recalculation only takes 1/100 of the second. Then with 20FPS, there is 5 mistakes per second, the 1/4 of data needs to be recalculated! And that recalculation takes 1/20 of s. And no one can say that error can’t appear then recalculating…
This explains the strange effect with extremely low FPS like 1-3. At this point there should be 1-3 FPS (not overclocked), after oveclocking, game simply freezes at this point. It takes 1-1/3 of the second to calculate it, and it gets error every 1/5 of second, this all goes into that this frame can never be calculated right = 0FPS! Sometimes this even goes into game crash.
Now lets go to something that is for sure, I can tell that the whole frame isn’t recalculated. Then you OC too much you see a hole in the wall this the part that had to be recalculated. All the objects are divided into such areas, and all this area is usually recalculated if needed. The one thing you should know, is that polygon count and FPS don’t effect number of areas. All this goes again, the more data there is, the more data in one area is, the bigger the possibility to get a miscalculation.
That’s all for now, I don’t have any more info and I hope that someone understood what I wanted to say.
Now lets go to letters, I received a bunch of them, and the most common question was how I measured it. So, 1st I used 3DMark2001 SE all game tests. How? Final 3DMark score is the sum of average FPS in all game tests. So I have the average FPS. Now lets go to min. I start “Fraps” and simply look for the lowest number. I also checked it in 3DMark03. And the 3rd is Sam 2, I simply looked for the lowest point, I didn’t have the average here. The easiest way is simply to play a game and look if the it works smooth.
All the tests gave same results, this shows that all this isn’t a mistake. You can write to me if you have anything to say. I try to reply to every email:
[email protected]
And I repeat once again: don't mind the numbers!
