GPU hardware chips [11.20]

 

 

Pixel shader

//

32 pS threads are a group, and each (one of 32) is responsible for one fragment.

Fragment requires 2x2 quad for processing.

Each thread can process data from up to four triangles.

Therefore, there will be a lot of waste: for example, the triangle generation is less than 8 pixels.

This feature always works on all N cards.

//

When the size of a triangle is smaller than one pixel, the bounding box test will be performed. The bounding box is the BB of the triangle, and the test will be performed on the surrounding pixels. If it is not at the intersection of any fragment center, discard.

That is to say, a triangle smaller than one pixel brings additional burden.

This feature always works on all N cards.

In frustum culling, we may also make a size test. If it is smaller than one pixel, it will be cull.

 

// Discard & clip

Commands after discard and clip will be skipped (on the PS3, only texture sampling behaviors will be skipped, and commands and Alu will still be executed ).

When the command is executed, it is a thread group (nv32, ATI may be a little different, but it is also dozens of) and clip will exit in advance. Otherwise, if one goes to the end, it will all go to the end.

 

 

 

Render target

• D24s8 and d24x8 are better under NV card g80, and D16 is a tragedy. It is corrected under the Fermi architecture.

 

 

Performance spec

You need to see:

• Architecture ---- The geforce8 series to the geforc4series are the g80 architecture, and the geforce4 series is the Fermi architecture. This Hardware Algorithm has changed and has the greatest impact on performance, including many changes to the algorithm pipeline.
• Specific parameters, number of cores, frequency, memory size, and frequency.

It is useless to check the product number. It is mainly used for marketing. gt240 is inferior to geforce8800gtx.