The L2 is cache used for things like large floating point or vecotor calcs, like updating the movement and postions in a game, rotating objects, moving shadow edges, shading, precalcs that are fed to the video card for finishing and display. The more that can be precalced the faster the slower card chip reacts. Cache can be 2-8X faster than main RAM, or more depending on CPU and RAM speeds. Complex vector calcs eat RAM or cache space. What you see in a game is a surface-enhanced wire grid motion sim. The underlying FRAME is claced to a large degree by teh CPU and the colors also calced. The card blends if told to right, to get shades that fade from color to color and simultaneously to depth.
So, bigger L2 cache ondie means the framework gets calced faster and sent to card faster than a calc that has to be pended in part to RAM. Only reason you do not have 1 GIG of L2 is simply the cost of mfring and verifying the goodness of that large a cache array. With these processors, both a small L1 and a large L2 are ondie-- future gen chips will also have an L3 cache ondie and that will be used as primarily a RAM buffer for RAM I\O. L1 is tiny,hyperfast cache that is very close to the same circuit set that forms the core calc pipes of a CPU. L2 is slower, still much faster than graphics RAM even, and is used as a buffer for I\O when no L3 or for calcs if not needed for I\O. When L3 is present, it is typically about the speed of graphics RAM. Main RAM is slowest typically, until you get into the PC3700 DDR which is closer to graphics RMA still in use on video cards that are not the fastest.
Now, if this is a remote only game server, the video card will be much less important and the cache on CPU even more important than local single user box would have, it will be handling many feeds of very complex 3D vector array calcs.
Comments
So, bigger L2 cache ondie means the framework gets calced faster and sent to card faster than a calc that has to be pended in part to RAM. Only reason you do not have 1 GIG of L2 is simply the cost of mfring and verifying the goodness of that large a cache array. With these processors, both a small L1 and a large L2 are ondie-- future gen chips will also have an L3 cache ondie and that will be used as primarily a RAM buffer for RAM I\O. L1 is tiny,hyperfast cache that is very close to the same circuit set that forms the core calc pipes of a CPU. L2 is slower, still much faster than graphics RAM even, and is used as a buffer for I\O when no L3 or for calcs if not needed for I\O. When L3 is present, it is typically about the speed of graphics RAM. Main RAM is slowest typically, until you get into the PC3700 DDR which is closer to graphics RMA still in use on video cards that are not the fastest.
Now, if this is a remote only game server, the video card will be much less important and the cache on CPU even more important than local single user box would have, it will be handling many feeds of very complex 3D vector array calcs.
That is the basics on caches for CPU.