Simplified, the more memory on a video card and the faster its graphics processing power with more pixel "pipelines", the better the image will be.
Each pixel "pipeline" represents one complete image that is ready to be displayed, in simplistic terms. The faster the GPU can assemble the data for the desired image, the faster the motion is for the graphics. The system's CPU has to decode all the data from program files and decide which bits are video, audio or other sprite/scene control bits.. as well as synch of sound to motion. The video bits must first be sent to the dedicated main memory or "aperture" set aside for the video card. The video data has to be then sent to the graphics bus for the video card to assemble in the correct resolution/color shadings desired, and then the audio portion must be sent to the sound card and properly timed so action/sounds are matched. The GPU does only about 1/4 the work of the CPU, but its work is very critical to the speed at which the images are displayed. The faster the data bus and the GPU's main 'engine' speed, the more realistic the images appear.
At an XGA nominal resolution of 1024x768 with an aspect ratio of 1.33, there are 786,432 pixels in each image.
At that same resolution with a refresh rate of 58 Khz. horizontal/60 Khz. vertical, that is beyond the limits of a graphics card that has only 4 Mb of memory and with a core speed of less than 100 Mhz. It would be better if the core speed were greater than 250 Mhz with a minimum of 128 Mb of memory on the card itself. The normal 1x AGP bus has a limit of 266 MB/s data transfer capacity. At 8x, that increases to 2133 MB/s data transfer capacity.
The memory on the video card may be as fast as main system memory, but the path to it is much shorter than that to the main system memory on the mobo.. so data transfer times are far greater. The main memory bus is shared with all data programming, but the video memory on the card itself has only one "gate" to pass for its task.. not two or three, as with main system memory. With no dedicated memory of its own for on-board video systems, the main memory must have the raw data assembled and then the CPU must then process it through the Northbridge memory controller for the GPU to render into an image. The CPU does double-duty with no add-on display card. With the add-on display card, there is only raw data formed in the aperture for the video, and that is then relayed to the video card to further process and render an image. The Northbridge memory controller simply passes data in a "pipeline" of its own.
A pixel is comprised of three color elements that can display up to 17 million-plus color ranges.
If you take a nominal resolution of 1024x768, as noted that is 786,432 pixels of one level. Give that single level a boost to 32-bit, and you wind up with an astounding 25 million-plus bits of data to be transferred in one second!