FP Blending / Filtering Benchmark?

[KimB]

I would like to see a 6600 GT test with larger texture sizes, to see if the same texture cache issues are in play here.

Anyway, it still seems like there was something significantly broken in these cards that was later fixed, either in just a more recent revision of the core, or the revision that made it into the PCI Express parts.

 

[DOGMA1138]

why dose the 9800 get better scores then the 6800 in large textures, but get powned by the 6800 in the small size ones?

 

[mczak]

why dose the 9800 get better scores then the 6800 in large textures, but get powned by the 6800 in the small size ones?

Well the 9800 should get owned - and as others have said, the low score of the 6800GT with large and float textures appears to be caused by some texture caching issues. I'd also be interested in results with texture sizes between 32x32 and 1024x1024 size, btw.

Stock 6800:
small 32-bit int texture w/ bilinear filtering, 32-bit int render target, no blending: 1949.3 Mpix/s
small 32-bit int texture w/ bilinear filtering, 64-bit FP render target, no blending: 2578.0 Mpix/s

Are you sure you're not forcing AF or something like that? Those results seem inexplicable in relation to each other:
It is faster with a 64bit FP render target???

 

[KimB]

why dose the 9800 get better scores then the 6800 in large textures, but get powned by the 6800 in the small size ones?

Well, the 9800 isn't doing any filtering or blending, so that's one reason right there why it should perform better (in comparison to its fillrate). But I think the main reason is just that there is something that's just plain not working properly in the NV40 that appears to have been fixed with either just the 6600 so far, or also with the other PCIe versions of the NV4x line.

 

[trinibwoy]

Stock 6800:
small 32-bit int texture w/ bilinear filtering, 32-bit int render target, no blending: 1949.3 Mpix/s
small 32-bit int texture w/ bilinear filtering, 64-bit FP render target, no blending: 2578.0 Mpix/s

Are you sure you're not forcing AF or something like that? Those results seem inexplicable in relation to each other:
It is faster with a 64bit FP render target???

Just did it again with the same results. No AA/no AF

small 32-bit int texture w/ bilinear filtering, 32-bit int render target, no blending:   1977.8 Mpix/s
small 32-bit int texture w/ bilinear filtering, 64-bit FP render target, no blending:   2554.8 Mpix/s