NV GF6系列架构分析（GPU gem2 读书笔记）

http://www.cnblogs.com/wangdaniu/archive/2006/02/20/334089.html

总线带宽：PCI Express×8, 66Mhz *32bits(4Bytes)*8 = 4 GB/s

目前计算机中的带宽：
GPU显存：550Mhz DDR ×256bits(32Bytes)×2=35GB/s
PCI Express ×16 ：8GB/s
CPU内存，800Mhz前端总线：6.4GB/s

显卡在PC体系中的位置：

CPU
                                                /\
                                                 |
                                            6.4GB/s
                                                 |
                                                \/
内存<=6.4GB/s or more=>北桥<=up to 8GB/s=>GPU=> to display
                                                /\                                    /\
                                                 |                                      |
                                                 |                                   up to 35GB/s
                                                \/                                     |
                                              南桥                                \/
                                                /\                                  显存
                                                 |
                                                \/
                                        其他设备

图形操作流程：

Vertex Data => Vertex  Shader×6 => Cull/Clip/Setup => Raster =>Pixel Shader ×16 =>Fragment Crossbar => Z/Blend => Back Buffer

Color Buffer => Pixel Shader (as texture)
Z Buffer=> Early Z Cull
Pixel Shader => Vertex Shader (maybe stream out)

顶点处理单元：
                         
顶点数据==> 32位浮点向量ALU×1
                         32位浮点数值ALU×1==>分支单元==>返回ALU
                         顶点纹理采样单元×1                        ==>或者生成Primitive==>culling/clipping==>光栅化

像素处理单元：

图元数据==>32位浮点ALU×1：执行一次乘加操作或者纹理采样==>
==>32位浮点ALU×1（一次加操作)==>分支单元==>返回ALU
                                                                                          ==>或者进入雾化ALU