why an nForce GPU is faster than GF2pro?

[BORG3577]

Hi I noticed something funny on my PC, maybe someone can illuminate me.
I have an nForce 2 MB with a 2400+ Sempron and an old GeForce 2 Pro.
Occasionally i play DOW andi get a frame rate of about 26. For some reason i removed my GeForce 2 and had to use the integrated nForce graphics and i got 36Fps!
i checked specs on internet for Geforce2 Pro it can do 800MPixels or 1600Mtextels
and the nForce can do 450MPixels or 900MTexels.

Is there some other thing that determines performance apart the pixels or texels per second?

 

[Rage187]

perhaps being integrated gives it an advantage.

 

[jiffylube1024]

Originally posted by: Rage187
perhaps being integrated gives it an advantage.

Come now, let's not start with that .

 

[Gstanfor]

The integrated nForce GPU features a 2 way memory crossbar.

The GeForce2 series lack a memory crossbar (first introduced with the GeForce 3) and as a result they have a tough time effectively harnessing all of their potential bandwidth.

The crossbar is what allows the GF4 MX series to perform as well as they do. The MX-460 performs better than a GF2 Ultra.

 

[Yanagi]

Did you double check settings for both graphic cards? Is the system overclocked, running without dividers. It might be that the integrated have more bandwidth to take from than the PRO hence the higher FPS.

 

[BORG3577]

could be the crossbar i remember its introduction a couple of years ago. Any way what amazed me is that the specs of Mpixels/second and Mtexels/second don't really matter!
wow the 460 outperforms the GF2 U!
according to specs
GF2U: 1000MP/s 2000MT/s
460MX: 600MP/s 1200MT/s

 

[Gstanfor]

Mpixels and Mtexels still do matter, but, if a GPU can't access enough of them per cycle they are useless.

GF2 without a crossbar has to acess memory in 128 bit chunks every cycle. If the data it needs is contained in only part of those bits, then the rest are wasted for that cycle.

With a memory crossbar you effectively split that 128 bit bus into 2 64 bit busses, which is good if the data you want to fetch fits into 64 bits (a lot of stuff does), so you can effectively do much more per cycle.