Bandwidth of NV30 and R300

[Sonic]

Would the NV30 benefit greatly from having a 256-bit bus instead of a 128-bit? And if so would that give it an advantage over the R300 in some cases? I'm just interested in seeing people's opinions on this subject since I've been out of the scene for some time. I've seen some graphics chip stuff that have 256-bit bus and it runs much better than any of the current arcade hardware setups. It's around 5 times as powerful as the Xbox board, and so far one fighting game is pushing a solid 50 million pps with a resolution of 1024*768.

Any thoughts if NV30 would perform better than R300 regularly if it had a 256-bit bus?

 

[Dave H]

Yes.

However, note that this is not as simple as just changing the PCB you stick the chip on. At the least you need to redesign the memory controller, and to get maximum benefit you need to resize caches, buffers, internal bus widths, etc. Still, I'd guess this is the change many of us are most expecting from NV35 (Nvidia's next top-end GPU after GFfx).

 

[DOOM III]

NV added some features into the chip to prevent the bandwidth from being the bottleneck,hence i think an upgrade from 128bit bus to 256bit won't give NV30 a big performance boost.

 

[jvd]

one site overclocked the ram 10% and only got a 1.5% increase in performance , so i doubt the 256bit buss will help

 

[Dave H]

You're probably referring to this quote from Anand's review:

Anand[/url]]We debunked this theory by increasing the FX's memory bandwidth by 10% (read: overclocked it to 1100MHz memory) and noted no more than a 0.2% increase in performance, so this test was clearly not saturating the memory bus of the GeForce FX.

The test in question is UT2k3 flyby at 1024x768, not the most bandwidth-limited test in the world. In fact, if you look at the results (it's a CPU scaling test), you'll see that they're completely linear (at least as well as can be divined from the graph) on the GFfx w.r.t. CPU clock speed. (The apparent discontinuity between 2200+ and 2400+ is because AMD changed the Quantispeed formula, essentially subtracting 100 points from 2400+ on.) So it's pretty safe to conclude that we're CPU limited here, and that GFfx's drivers are taking up more CPU time than the 9700's.