Why not integrated DRAM/chunking (yet)?
In 0.25u vlsi process, 1.2-1.5M logic transistors == 1MB of IDRAM
IDRAM is about 2X cost of pure DRAM
compare transistor budget of approximately 7M:
- 4-4.6M transistors graphics engine + 2MB IDRAM frame buffer
- 1-2.2M transistors graphics engine + 4 MB IDRAM frame buffer
- 7M transistors graphics engine + 8 MB or 16MB SDRAM frame buffer
fill rate in all cases > 180mpix/sec (not the bottleneck!)
local texture storage limited in IDRAM case
triangle rate limited in IDRAM case, if chunking
- CPU must create chunks (1.2-2X overhead for processing chunk data)
- CPU must manage chunk data (2X AGP overhead for extra data copies)
- Graphics must process chunks (ɭM transistors for chunk creation HW)
- Graphics must store chunks locally (competes with FB, limits triangles/frame)