Testiculus Giganticus
Newcomer
8)
GF FX performance?
- Thread starter sir doris
- Start date
Joe DeFuria said:
How can you do 8x "ordered grid" in a way that would share the same lack of quality as 4x ordered grid such that it would be inferior to 4x rotated grid?
OK, so there is the obvious idiocies like 4x OGMS + 4x OGSS. I hope that such redundancy can be agreed to be rather unlikely and would require that 8x mode be 100% pure marketing drivel and nVidia is insane enough to think they can get away with it. Dismissing that, what is the worst it could be?
It seems likely to me something like 2x2 OGSS + 2x RG MS, given what we already have. Besides being slower, how bad is this compared to 4x RGMS?
EDIT: given the confusion TG introduces, what would simple sub-pixel offset make a two 4x grids look like? What if the offset was diagnol? Wouldn't this effectively give coverage like rotated grid, though the sampling pattern is ordered grid?
Joe DeFuria
Legend
Ask nVidia.
I (and everyone else) really donen't know...which is why this thread is getting particularly long.Givin nVidia's "pipeline architecture tuned to 4 samples per pixel", I would be inclined to disagree. I think they would utilize 4X multisampling where possible. That would seem maximize performance of an "8 Sample" algorithm, (2X1 supersample + 4X multisample, both ordered grid). Compared to your 2x2 Supersample + 2X multisample, it should be faster, at the expense of image quality
In any case, the problem is...all of our suggestions seem to have some inconsistency when trying to reconcile with the interviewer, and what we know nVidia has tradtionally done with their techniques and naming conventions.
For example,one would think that my suggestion would be called "8XS", not 8X....
Joe DeFuria said:
Ask nVidia.
OK, lets work at it from a different angle. I think nVidia does not implement combined supersample and multisampling in OpenGL (hence no 4xS), due to some implementation and OpenGL specification detail. 8x will be implemented in OpenGL, 6xS will not.
Crazy idea one: z-buffer 8x the size. That's a decent chunk of memory, but don't you think it is feasible? Seems rather brutish though, but I think it would look pretty good.
Crazy idea two: 2 renders with offset sample positions, both ordered grid to fit the interviewees comments, combined later (for "free", in hardware?). If it could do this, it could possibly do some other things that would support the AA hype comments. It also seems to fit things I seem to vaguely recall about the "mysterious 3dfx tech" suite.
Well, if 2x RGMS + 2x1 OGSS is called 4XS, your suggestion sounds like 6XS to me.
When I mentioned worst I really meant that is the worst I think it could be...looks like you couldn't think of anything worse either.
Joe DeFuria
Legend
Heh...problem is, the interview clearly stated that 8X does combine multi-sample and supersample.
Like I said...inconsistencies all over the place. We're likely not to get anywhere without further clarifcation from nVidia...Not IMO:
1) Render to 2x1 Supersample buffer...
2) Apply 2x2 multisample to each of those supersamples.
So the result is 8 samples. (2 supersamples, times 4 multisamples). That is consistent with the naming of nVidia's 4xs. (2 supersamples times 2 multisamples).
With all the confusion going on here, ir could be worse...nam_ng could be in on this conversation...
If the xS means skewed then that makes some sense as there is no 8xs its just 8x. 4xOGMS with 1x2OGSS would give an ordered grid 8x pattern like ati have said.
For 6xS it could be 2xRGMS with 1x3OGSS. That would give a skewed grid and theres probably a few other ways to do it too.
The only problem with the above is that for effective near horizontal and veritcal edge AA 8x is barely any better than 4xS and 6xS would be better than both. Also the 9700s 4xAA would be better than all these.
Hmm speculation
For 6xS it could be 2xRGMS with 1x3OGSS. That would give a skewed grid and theres probably a few other ways to do it too.
The only problem with the above is that for effective near horizontal and veritcal edge AA 8x is barely any better than 4xS and 6xS would be better than both. Also the 9700s 4xAA would be better than all these.
Hmm speculation
Joe DeFuria
Legend
Hmmm...that I didn't think of.
Actually, not too much of a problem...it would just be a bit odd. In the sense that 8X could actually turn out to be be "faster, but lower quality" than 6xS. Counter intutive because 8>6.
Man...I could see web reviews and product comparisons really f*cking things up if this turns out to be the case.
megadrive0088
Regular
*waits to see if NV35 has a better anti aliasing unit*