8800GTX Shadermark results
- Thread starter Geeforcer
- Start date
- Status
I wonder. Does it mean that nVidia has implemented an on-the-fly trilinear filtering technique? Or are the texture units just meant to perform two bilinear filters per clock? The first could be good for image quality/performance ratios (though the image quality will be somewhat lower than "true" trilinear, it should be much better than brilinear). The second would be a good optimization for anisotropic filtering, though nobody should expect trilinear to be free with AF enabled in that case.
Edit:
There is a third, much worse option: trilinear filtering isn't actually being enabled. I hope this isn't the case, but it is possible.
Edit:
There is a third, much worse option: trilinear filtering isn't actually being enabled. I hope this isn't the case, but it is possible.
The other possibility is simply that the rate at which you can filter is uncoupled from the rate at which you can calculate texture addresses. (e.g. you can only calculate N texel addresses per cycle, but have 2N filtering units, then trilinear is "free" in the sense that if you don't enable it, you have bilinear units potentially going to waste)
Yeah, that's kind of along the lines of what I was thinking. But now that you say it like that, it makes some sense that this might possibly be true free trilinear. Anyway, I'm really excited about this architecture...definitely looks like lots of interesting changes!
I still haven't fully understood the texturing unit organization on G80, but if the first possibility should be real, remind me to give you a virtual kick in the butt for past conversations related to single cycle trilinear we had.
Uhmmmm LOL I doubt that
Ah, c'mon. Dogs and cats living together, frogs from the sky, John Reynolds chaining himself to the front doors of NV headquarters. . . . just thinking "not so much" there, Chal. (Well, okay, maybe JR would chain David Kirk to the doors. . .but there'd be chains and doors!)
Unless a bug has slipped into a driver, all drivers ARE usually set to a default "0" LOD bias. At least up to G7x you COULD NOT alter the MIPmap LOD bias values unless you used a 3rd party application like Rivatuner.
There's of course also the clamp function (which should be gone on G80 since I don't see it's use anymore) which forces the application to stay on default 0 despite the application calling for negative LOD bias, but it's entirely irrelevant here too.
OK, folks!
It looks as ArchMark likes to throw out big bunch of numbers, so I've decided to put some order in the field and here is the result, with few gathered reference scores aside to the star of the day [G80]:
p.s.: it was a real pain in the a** to collect and organize all the data in the chart, but I hope it worth.
It looks as ArchMark likes to throw out big bunch of numbers, so I've decided to put some order in the field and here is the result, with few gathered reference scores aside to the star of the day [G80]:
p.s.: it was a real pain in the a** to collect and organize all the data in the chart, but I hope it worth.
Last edited by a moderator:
OK, folks!
It looks as ArchMark likes to throw out big bunch of numbers, so I've decided to put some order in the field and here is the result, with few gathered reference scores aside to the star of the day [G80]:
p.s.: it was a real pain in the a** to collect and organize all the data in the chart, but I hope it worth.
thanks! G80 looks phenomenal.
You may like to note that some texture caches my report differently dependant on whether there's and L1 and L2 cache available.p.s.: it was a real pain in the a** to collect and organize all the data in the chart, but I hope it worth.
Thats not true for some texture formats on (at least) NV4x.
- Status