Bitboys over fillrate en memory bandwidth

Mika Tuomi van Bitboys Oy (onze Finse Glaze3D vrienden) heeft bij Voodoo Extreme een "Understanding fillrate and memory types" artikeltje gepost, als follow-up van z'n opmerkingen over de beperkte fillrate van GeForce 256 (zie deze nieuwsposting):

To calculate theoretical maximum fillrate you need to know available memory bandwidth, pixel size (16/32bit), Z value size (16/32bit) and texel size.

When a 3D accelerator renders a triangle, it splits the triangle to single pixels. Normally 3D accelerator does the following sequence for each pixel:

- Read old Z value from Z buffer
- Write new Z value to Z buffer
- Read Texel(s)
- Write new Pixel to frame buffer

From this sequence we can calculate maximum pixel fillrate for a given bandwidth.

In the first calculation example we have 32bit Z buffer, 32bit frame buffer, and we read one new 32bit texel for each pixel. 32bits is 4 bytes (32bits / 8 bits). Normally texels are read from onchip texture cache, so this one texture read per pixel is just one possible behavior of texturing. This example can be considered very high quality rendering.

- Z read (4 bytes)
- Z write (4 bytes)
- Texture read (4 bytes)
- Color write (4 bytes)

This sums up to a 16 bytes that have to go through the memory system for each pixel. Now we can take our first memory system example (128bit @ 166MHz SDRAM) and see how many of these pixels go through it. 2.656GigaBytes/sec / 16Bytes = 166MegaPixels/sec.

Nu zou ik wel eens willen weten hoe dit bij de Glaze3D werkt.