NV40 has 1 TMU? FillRate is not same. (single vs multi)

[zero12]

I got it DarkCrow.

I know that R3x0(8/1) has same fillrate between single-texturing and multi-texturing. But NV3x has 4pipes and 2tmu, so it has difference fillrate among single and multi.

This picture shows NV40 has different fillrate.

 

[DemoCoder]

It has one TMU. The difference could have something to do with the efficiency of single texturing vs multitexturing when using alpha blending.

 

[zero12]

Thanks DC.. I have some confusion.

 

[GameCat]

I'm not sure about 3DMark03, but in 01 fillrate tests were mostly bandwidth limited so they didn't actually measure fillrate. If they're still doing alpha blending to test fillrate the NV40 is probably bandwidth limited in the single texturing tests.

 

[Neeyik]

Read Dave's preview - the single texture test is bandwidth limited hence why the fill rate delta. It's more pronounced on the NV30 because one is looking at 16 pipes trying to do this compared to the 8 on the R3xx. It's slightly odd that the test is called single texture when it's really applying 64 over 64 surfaces, blending the whole lot and then producing one frame.

 

[cho]

here is my fillrate test that use the "fillratebenchmark2004 0.92b" :

http://www.gzeasy.com/ours/benchmark/FillrateBenchmark_V092_b.zip

 

[{Sniping}Waste]

Sorry for the OT but looks like something wrong with the Vertex shader. Its only a small improvment over the NV3X. Look at the Vertex shader test, ragtroll, and trolls lair. Ragtroll and trolls lair is very in vertex shaders if I remember right.

 

[Neeyik]

Sorry for the OT but looks like something wrong with the Vertex shader. Its only a small improvment over the NV3X. Look at the Vertex shader test, ragtroll, and trolls lair. Ragtroll and trolls lair is very in vertex shaders if I remember right.

Nothing wrong as such - a 5950 Ultra will typically score around 19 fps in the VS test, so the increase is pretty in line. Rag Troll contains a fair bit of CPU work too, so it's not entirely VS limited and Trolls' Lair is only VS-bound is certain scenes.

 

[stevem]

Cho, could you post D32 results too?

 

[Mintmaster]

I'll second that motion. I think ATI disables Z-compression when 16-bit Z is used, and NVidia might do the same. That would explain the 32-bit colour + 16-bit Z fillrate being somewhat low. I would expect near perfect Z-compression in a fillrate test.

I'm also surprised at the 16-bit alpha blending performance. Anyone think NVidia only put 8 blending units on the card? It's very reasonable, though, for 32-bit rendering. I wonder how many FP blending units there are. It would make sense (from a bandwidth point of view) for there to be only 4, but hopefully NV40 can still FP blend more than 4 pixels when rendering into a single or double channel formats (like D3DFMT_R16F).

 

[DemoCoder]

Mint, you're very observant.

 

[cho]

the 60.72 does not provide D3DFMT_R16F flag yet .

 

[StealthHawk]

Sorry for the OT but looks like something wrong with the Vertex shader. Its only a small improvment over the NV3X. Look at the Vertex shader test, ragtroll, and trolls lair. Ragtroll and trolls lair is very in vertex shaders if I remember right.

You should probably also take into consideration the fact that NV40 is running a Futuremark approved driver and NV3x isn't