专业人士说64位宽(转贴)

mikesun

引用自显卡位宽知识:
     显存位宽是显存在一个时钟周期内所能传送数据的位数，位数越大则瞬间所能传输的数据量越大，这是显存的重要参数之一。目前市场上的显存位宽有64位、128位和256位三种，人们习惯上叫的64位显卡、128位显卡和256位显卡就是指其相应的显存位宽。显存位宽越高，性能越好价格也就越高，因此256位宽的显存更多应用于高端显卡，而主流显卡基本都采用128位显存。
　　大家知道显存带宽＝显存频率X显存位宽/8，那么在显存频率相当的情况下，显存位宽将决定显存带宽的大小。比如说同样显存频率为500MHz的128位和256位显存，那么它俩的显存带宽将分别为：128位＝500MHz*128∕8=8GB/s，而256位＝500MHz*256∕8=16GB/s，是128位的2倍，可见显存位宽在显存数据中的重要性。


引用自 [Microsoft DirectX 9.0] 关于图形资源模型部分说明:
Resource Properties
All resources share the following properties.
Usage. The way a resource is used, for example, as a texture or a render target.
Format. The format of the data, for example, the pixel format of a 2-D surface.
Pool. The type of memory where the resource is allocated.
Type. The type of resource, for example, a vertex buffer or render target.
Resource uses are enforced. An application that will use a resource in a certain operation must specify that operation at resource-creation time. For a list of the usage constants defined for resources, see D3DUSAGE.

The D3DUSAGE_RTPATCHES, D3DUSAGE_NPATCHES, and D3DUSAGE_POINTS constants indicate to the driver that the data in these buffers is likely to be used for triangular or grid patches, N-patches, or point sprites, respectively. These flags are provided in case the hardware cannot perform these operations without host processing. Therefore, the driver will want to allocate these surfaces in system memory so that the CPU can access them. If the driver can perform these operations entirely in hardware, it can allocate these surfaces in video or accelerated graphics port (AGP) memory to avoid a host copy and improve performance at least twofold. Note that the information provided by these flags is not absolutely required. A driver can detect that such operations are being performed on the data, and it will move the buffer back to system memory for subsequent frames.
For details about the usage flags and how they relate to specific resources, see the reference pages on the individual resource creation methods.


一般来说， 64位通道线可以实现4GB/秒的传输速度,  但受到[物理内存寻址+读],[主板传输],[显卡显存寻址+写]速度的影响, 一般无法达到 4G.

而且根据DX9.0中的字句:
the driver will want to allocate these surfaces in system memory so that the CPU can access them. If the driver can perform these operations entirely in hardware, it can allocate these surfaces in video or accelerated graphics port (AGP) memory to avoid a host copy and improve performance at least twofold
设备将试图在系统物理内存中分配一段内存用于给CPU访问.  但如果硬件能够完全的实现这些显示操作, 则设备将直接在显存中分配内存来显示图层,  这样可以避免物理内存到显存中的拷贝操作,来提高至少2倍的性能.


结论:
      象X1600显卡,可以将大量复杂的操作,比如物体渲染,缩放,旋转,透明度等3D中几乎复杂的操作,都直接放在显存中来进行处理.  其传输性能的瓶井将在与硬盘读取速度!   但硬盘读取速度要想达到1G, 那估计还要N年, 大家参考下硬盘对拷一个2G左右的游戏就知道了,硬盘的读取速度顶多也有50M.

     而目前我们知道, 操作系统\软件\硬件这几年进行了32位到64位的革命性过度， 编程再菜,也知道结构体要按系统支持的字对齐吧? 比如32位对齐, 64位对齐,来适应CPU,内存对单字处理的效率.  64位系统能够一次对一个64位单元进行完美的处理, 如果上32位的显卡, 就必须把一个64位字进行两次传输. 而128位呢?  目前128位的系统\软件\硬件还没出生~~~

所以说, 目前128位显卡, 256位显卡, 基本只是嚎头...   
你开个小厂, 生产部每年生产100吨货物,  你运输部一年可以运1000吨货...

啥也别说了, 直接看3D得分, 只在位宽上有区别的显卡, 位宽64位以上的,  实际得分将是一模一样的......