Overclocking: Core vs. Memory

There's reviews everywhere comparing graphics cards to each other, but I've never seen one that showed the gains of overclocking any single one (I won't even mention detailed overclocking comparisons of several cards). The closest to that are the few that have a "before & after" performance comparison for a single card or sometimes multiple cards, but that's not very much information about how changing the clock varies the performance. As you read this (well, for the one or two people who actually will read this), keep in mind that the entire article deals with the graphics card only, and not anything else!

I decided to push "Benchmark" about 50 times while studying SAT words and find out for myself: what effect on performance will changing the core speed have if memory speed is held constant, what effect will changing the memory speed have if the core speed is held constant, and finally, what effect (you guessed it) will changing both of them have?

If you aren't familiar with the term overclock, please don't try to find out through experimentation. You may end up with undesirable results such as fire; explosions; toxic fumes from melted plastic, scorched carpet, singed fake wood desk, vaporized metal or all of the above; or more realistically, crashing, instability, and permanent damage to your hardware (actually, this includes a small risk of melting and a large risk of totally destroying a part or parts).

If you're really curious, which I'm not expecting, then feel free to ask me to clarify anything.

Conventions/Definitions

The conventions and definitions used throughout this article include:

SDRAM (or just RAM), memory, and DDR (Double Data Rate) are used interchangeably. All memory or RAM referenced in this article is DDR SDRAM. DDR means that the RAM transfers on both the leading and falling edges of the clock cycle. This means an effective doubling in data transfer rate (hence the name). So if DDR SDRAM operates at 166MHz, then its effective speed is 333MHz although it is not actually running at this speed. The effective speed (the larger number) is the one that is always used in this article. The memory mentioned in this example will therefore be referred to as DDR333 or 333MHz DDR (although technically this is incorrect, as explained, it is customarily done this way), not 166MHz DDR (which is the more correct way of saying it).
Core - in this article, core refers to the GPU or graphics processing unit. It is the brains of the graphics card, analogous to a CPU in a computer system.
Memory - when not in some other context, it refers to the DDR on the video card; may also mean system memory.
Clock, speed - These are used synonymously to mean, well, the clock speed of a part. It basically measures the how fast something is running. For example, a Pentium 4 CPU may operate at a clock or a speed of 2800 MHz.
Performance, 3DMark Score - Used interchangeably. I'm using the benchmark as a direct measure of performance, which is not all that bad, considering that it is a series of game tests and is not totally synthetic (but of course not totally real either).
Artifacts - incorrectly rendered (drawn) elements that look funny (well, sometimes not so funny) and are not supposed to be there.
Don't worry about my multiplication. If it seems off by a tenth or two, there isn't actually a problem. It's roundoff error. For example, 166.7*11 = 1833.7 (see Hardware just below), but it's actually 166.66666..., so this is where the 1833.33 comes from. Sig figs are somewhat observed here. I didn't really make an effort to, but I kind of do that automatically anyways.

Setup

Hardware

ASUS A7N8X Deluxe (nForce2 chipset), PCB v1.04
AMD Athlon XP (Barton) 2500+ running at 198.7MHz FSB x 11.0 = 2185.3 MHz.
Default clock is 166.7MHz FSB x 11.0 = 1833.3 MHz.
2x256MB DDR400 (PC3200) 2-3-3-6 Corsair XMS memory running at 198.7MHz x 2 = 397.3 MHz.
XFX Geforce4 Ti4200 128MB 8x AGP with VIVO.
Default clock is 250MHz core and 513MHz memory.

Software

RivaTuner RC 14 (yes, I'm using outdated software because I'm lazy)
3DMark 2001SE

Procedure

Before starting, I attempted to increase my FSB as far as possible to reduce any memory bandwidth problems causing video card performance issues. I think that this has worked pretty well; it seems that the main factor limiting my graphics performance is the video card, and if the CPU was faster it would help a little too, though not as much as a faster video card. You may say "DUH!" to this last comment, but you can't always say that a faster video card results in better performance. This is another subject that may be covered at another time. If you want to know, you can always ask me.

Ok, so I grabbed RivaTuner and entered a lot of values for the clock speeds:

RivaTuner setup

Then I set 3DMark to all defaults except for the tests to be run, in order to save time. Running all of the tests results in the "all defaults" 3DMark score, which is a little different from what you get when only the first four are run.

3DMark setup

Then I just sat there and pushed Benchmark 50 times while changing around some clock values.

[ Setup | Results ]