I think there is quite bit of variation in the memory cost associated with each add-in board architecture. For example, if a graphics chip operates with a 64-bit memory bus it can use commodity (DIMM-type) 4Mx16 SDRAM (64 Mbit, 8 MByte a piece) with a 32MB framebuffer.
However, with a true 128-bit memory datapath, such as used by most Rage 128's, 2Mx32 SDRAMs are required for a 32 MB configuration which are likely to be priced higher. Moreover, for a 16MB configuration, the required 1Mx64 SDRAMs do not exist forcing the use of four times as many chips (4 Mbit, 2 MByte each), resulting in a significant cost disadvantage for a 128-bit vs. a 64-bit architecture. TDFX's solution in its new 16MB Velocity business board (128-bit memory) is to use SGRAMs to reduce the number of memory chips required (probably not ideal).
I think this is one of reasons for ATI pushing 32MB so strongly -- memory cost is much more competitive at 32MB compared to 16MB, since ATI cannot afford the performance hit of a 64-bit memory bus. Interestingly, NVIDIA does have a downgraded TNT2 product called "Model 64" (similar to Vanta) with a 64-bit bus which is targeted at add-in cards, reducing memory cost by allowing the use of 64 Mbit SDRAMs.
Of course, at the cutting edge of memory clock speeds (say, 160 MHz+) commodity-type memory parts are not usable, although ATI's memory clock speeds are probably still low enough to use commodity chips across the board. |
