Ok i'll try to give a kind of explanation .....
gpu redering can be differentiated into 2d and 3d mode ... in 2d images there's not much to render ... so the cpu takes the load and shifts only the display load to the card .... now the actual use of the card starts only in the 3d mode .. here the cpu decide to render only things related to logic/physics .... the graphics card renders all the frames,testures,antialias ,,,,etc ....the workload is shifted more to the gpu than the cpu ....thus rendring is completely shifted to the gpu ... while the cpu only takes care of the logic of the program ...
The presence of graphice card increase picture quality ...and fps in games even at high settings ....
Now the cpu interaction with the system ram is very slow ... i mean slow to render highly detailed frames ingame ....so the graphics card comes with its own ram which are much faster ....for ex ....the ddr400 ram modules actuallly operate at 200MHz with a 64bit interface .. a 6800GT comes with its onw ram modules that operates at 1000Mhz at 256bit interface ... so the loaded textures,rendering details can be transferred to the gpu at quite a high speed ... current gpu's operate at 500Mhz ... which is very efficient in handling the high amount of textures loaded into the gpu memory ...moreover due to the absentof any kind of a bridge the gpu-ram interaction is very high ...
The main reason behind the cost is the quality of fabrication ,the type of ram modules used in the card , and the class it belong to ....cards that cost around 5k comes with 64bit/128bit mem interface oprating at around 250/400 speeds .... cards like the 6600GT comes in at 12k now-a-days they have 128bit interface and operates at 500/1000 ... high end card comes with 256bit mem interface ..radeons oprating at 550/1150 ,, nvidia's monsters at 400/1200
128mb ram of the card means that the card can load 128mb of textures at once ....so the more the card ram the better the rendering ... but be warned this is not a full proof logic .... unless the gpu can render all the loaded textures at a fast rate there will be no significant boost in performance ... so go for 256mb card only if u intend to buy a mid/high end card ....most current games load around 128mb textures (except doom3) .. so a 128mb FX5700 or a 128mb 9600XT would be enough to run those games smoothly ....
No system ram and graphics cards ram are not the same .... the system cant share the gpu ram ...but the reverse is true in case of newer cards ....newer cards with turbo cache technology are able to share system ram over the pcixpress bus thus increasing the gpu ram ... but be warned system ram operates at much lower frequency than the gpu's ram so sharing system ram wouldn't be a good leap towards better performance .....