PDA

View Full Version : Pixel-Shader performance of low-end GPUs


Linuxhippy
10-08-07, 12:00 PM
Hello,

I am currently converting a software-only 2D graphics library to use fragment-shaders for at least the most expensive operations (it will still be mixed CPU/GPU).

What I wonder is how do newer Low-End boards perform compared to the (old&ugly&slow) FX5200 for pixel-shaders?
Could the following table be correct, and how would you rate pixel-shader performance for ??-marked cards compared to the FX5200/NV34:

FX5200: 100%
GF6200: 150%
GF7300: 300%
GF8400: ???%

X300: 150% (??)
X1300: ???%
HD2400: ???%

Of course I don't expect precise numbers, just a vague approximation what I can expect from those cards.

Thank you in advance, lg Clemens

walterman
10-08-07, 01:38 PM
Of course I don't expect precise numbers, just a vague approximation what I can expect from those cards.

I do not remember the specs of those low end parts, but, i think that you could use as approximation, the number of pipes * core freq.

Also, since the NV40, the chip can do up to 8 ops per pixel (shaderUnit1|Texture + shaderUnit2 on NV40 vs Texture + Shader on NV30).

In the G70, they added 2 mini ALUs to each Shader Unit, and the chip was able to do 8 MADD per clock cycle, vs 4 on the NV40.

The architecture of the G80 is a master piece. Basically each Stream Processor is a IEEE754 ALU, capable of 1 MADD per clock cycle. The efficiency of the G80 is higher cause on a scalar design, you can have more ALUs working at same time, than on a vectorial design (in which you need to assign the same op to the array of ALUs).

So, basically, for your numbers, i think that you can do that pipes * freq product, and then you can multiply by 2 if it's a NV40, or by 4 if it's a G70, or by 4...8 if it's the G80.