Results - Everyone's favorite
Putting everything written here into practice, we can see what types of gains can be made from overclocking. Please remember, these gains will differ from person to person, so if I score an increase way more/less than yours, don't be concerned.
The first thing I should mention is my computer specs, which are:
AMD Athlon XP 1600+
Thermaltake Volcano 6Cu+
Soltek 75DRV-2 Mobo
256MB Kingmax PC2100 DDR RAM
Hercules GeForce3 Ti200
Seagate Barracuda IV 40GB HDD
SB:Live 5.1 DE
A Laser Mouse!
Win 98SE (I still can't part with it .)
And other stuff that's irrelevant.
The drivers that I used in these tests were Detonator v23.11. My CPU was overclocked to 1478MHz, which isn't that much but it allowed me to raise the FSB to 140MHz. This gives slightly more bandwidth to the processor, which is important, as the bandwidth is the limiting factor for AMD chips at the present moment. Now for results.
The best program to test overclocking gains is probably 3Dmark2001SE, which tests basically all video card functions, including new DirectX 8.1 functions (which only Radeon 8500 and GeForce3/4's support). This doesn't test real-world gains like Quake III and Serious Sam, but it gives an excellent overview of general gains.
The first thing to note here is the different memory speeds listed. The top speed, 235/495, was completely stable, but above that proved to be erratic at times, and I thus decided to keep 235/495 as the top speed. Remember that is only 5MHz core and memory, from the GeForce3 Ti500 speeds, so it still gives an excellent representation of what a Ti500 does.
Overall, I managed to gain approximately 1000 3DMarks from the video card by overclocking. This came through a core increase of 60MHz and a Memory increase of 95MHz. When looking at the performance increases from overclocking the memory, you can equate it back to an equation which gives the maximum possible bandwidth the card can provide. The maximum possible bandwidth is, like the name suggests, the maximum amount of data the cards memory can move in a prescribed time, in this case, and most cases, one second. The equation is:
Memory Bandwidth = Memory bus width * memory speed / 8
The memory speed here is the actual speed it is running at, which means 200MHz DDR RAM would be included in the formula as 400MHz (since that is the speed it runs at). The memory bus width is measured in bits. The GeForce3's bus is 128-bit DDR. By adding the numbers in, you get:
128 * 400 / 8 = 6400MB/s
And that is the memory bandwidth of the card when running at stock speed. 6400MB/s looks like a VERY impressive number, but in reality it isn't brilliant. The GeForce4 Ti4600 has much a higher 10400MB/s, which is part of the reason it is so much faster. In most cases, increasing the memory bandwidth will increase the performance of the card. To increase the bandwidth, you raise the memory speed. By moving the memory speed a modest 20MHz, to 420MHz, we now have a memory bandwidth of 6720MB/s. This increase certainly doesn't set the world on fire, but it will help slightly. Moving the memory to the GeForce3 default speed we get a bandwidth of 7360MB/s. This is now 1000MB/s more, which is quite significant and means 1000MB more can be moved each second.
One thing to note is that this card, along with most good Ti200's, comes with 4ns RAM. 4ns RAM is actually rated to run at 500MHz. This can be seen using the very simple formula:
RAM Speed = 1000 / ns
So: RAM speed = 1000 / 4ns =500MHz (the speed at which the GeForce3 Ti500's memory runs, but that card uses 3.8ns memory). If we could run the memory at 500MHz, we could boost our memory bandwidth significantly.
Bandwidth = 128 * 500 / 8 = 8000MB/s
8000MB/s is a very handy 1600MB/s increase over the memory bandwidth of the card at its stock speed. Unfortunately, having 4ns RAM doesn't guarantee you being able to run at 500MHz, but if you can't get to that speed, you can usually get close (~480MHz is common, which is still faster than the original GF3, which had 460MHz memory).
There are also formula to calculate things like fill rates, triangles per second, pixels per second and so on. These are similar to the ones above and honestly don't mean anything. For example, by working out your fill rate and a few other numbers, you can work out your maximum possible FPS in a certain resolution. You will find you get very large numbers, which obviously are impossible to obtain in normal usage. Thus I have not added irrelevant numbers like that. I did include the memory bandwidth however, because it gives a decent indication of what overclocking does, and because memory overclocking does more for performance than core overclocking.
Another interesting point that shows in these tests is that a certain percentage increase in speed doesn't make the scores increase by that same percentage. At the default GeForce3 Ti200 speed of 175/400, I scored around 6600 3DMarks. I then raised the memory and core speed by 10% (just over 10% in the cores case), and recorded a score of 6800 3DMarks, which certainly isn't a 10% increase in speed. In fact, it equates to a lowly 3% increase in speed.
If we look at the percentage increase we get from overclocking the card 10% on just the memory, we can confirm that overclocking by a certain percentage really doesn't equate to that percent in performance gains. At 175/420, a 10% increase on the memory, I scored 6678 3DMarks. This is a measly 52 point increase. This pitiful increase may be due to benchmarking inconsistencies, but I am tending to believe not as other, large increases in speed have also given small 3DMark gains.
So as we can see, a certain percentage increase in speed certainly doesn't mean you get that same increase in performance. I wish it did, but nevertheless a 23.75% increase in memory speed and a 34% increase in core speed did give a 1000 point increase in 3DMark2001SE (about 15%).