Evildeus said:
It's you that aren't looking at what i said. You are saying that because it's broken on NV30 then Nvidia must be tweaking/cheating on NV35.
Umm...I'm saying the NV30 has floating point processing issues, and the info that we have for the NV35 indicates the same thing (twice as fast fp32 than the NV30 is still slow in comparison to the R3xx). I'm saying that drivers for the NV30 have seemed to use integer processing resources and shortcuts to make up for that, and that the NV35 would have the same need (fp32 performance doubling only brings it up to the fp16 performance of NV30, which is still not competitive with the fp24 performance of the R3xx).
Your statement above is based on a logical construct that is completely in the abstract, my reasoning is not.
I say, why not think that's it's broken/tweaked on the NV30 and the normal path on NV35? Half empty or half full?
Please, actually give some thought to your comments and relating them to my statements. To restate some observations, with interpretation spelled out in detail since it seems required:
You linked some screenshots.
One set was a teapot, with 3 shots, with no performance results that I see (this is something that if you see differently, you can tell me to correct me).
One shot was labelled 44.03 NV35/R3xx/REFRAST, one was labelled 43.45 NV30, and one 44.03 NV30. This does seem to tell us that the NV30 has different issues with 44.03 than 43.45, and the simplest explanation (without disregarding the actual hardware) seems to be that the default behavior for the NV30 is different for the 44.03 drivers. Since it has faster fp16 performance than fp32, it makes sense that arbitrarily promoting partial precision when full precision was expected is the explanation, and that the 44.03 demonstrably behaves differently for standar PS 2 between the NV30 and NV35.
However, that is standard PS 2.0, and unless you are being rather selective in what you are considering, it has been evident that for the higher performing 3dmark03 drivers, the PS 2.0 and GT 4 tests were not using full precision PS 2.0, but rather special case behavior. Which brings us to the next set of screenshots...
One set was of the GT 4 3dmark03 scene, with 2 shots, and no performance results.
One shot was "5800 43.45". The other was labelled "9800,GF FX 43.51+". It is known that the 43.51 showed the bright spots people pointed out as missing in the 43.45 for the NV30. It is also known, from screenshots I've seen myself, that other changes are evident for the NV30, and there is further indication that there other issues used to achieve its performance that will be revealed in the near future.
Also, performance results were shown for PS 2.0, but no screenshots were provided.
Are we tracking so far? I'll get back to this in a bit.
Moreover, you are disregarding the evaluations because it's not going your way of thinking.
If you'd bother to address my provided "way of thinking" coherently, I'd maybe view this statement as something other than saying your way of thinking is just as valid because you personally hold it, nevermind the things it seems to ignore. I just can't view that type of statement as useful at all.
Perhaps it the same for me, but contrary to you, the reviewers seems to think as me.
I don't trust others to think for me, and I'd actually like screenshots to atleast begin to be able to judge
for myself. This is an issue you disagree with, that is fine. But you seem to require that I point out details in answer to a complete absence of such from yourself.
In any case, I direct you again to note that the shot with the "fixed" GT 4 image is labelled as
GF FX 43.51+, the only place where 5900/5800 suffixes aren't used in the screenshots. FYI, though this seems obvious, GF FX=both the NV30 and NV35, and the site seems to be proposing that 43.51 (and higher) fixed this issue with precision (as some sites do).
The problem here is that the GF FX for 43.51 is labelled as equivalent for the R3xx for GT 4, when I directly know, based on my own independent evaluation of full screen shots of 43.51 in comparison, that I have not observed this to be true. You maintain that because the NV35 displayed properly for the teapot, that this indicates that the PS 2.0 and GT 4 results are displaying the same quality. What is clear from the selection of screenshots is that the only reason for that is based on the decision the reviewer already made for you, not information they provided for your evaluation.
The site seems to have made an assumption based on the brightness of the sky that is directly contradicting my own evaluation. This does not sit well with me for allowing their evaluation to substitute my own concerns. I think that if it sits well for you ignores things that are evident to someone who is thinking for themself, but rather than try and persist in dissuading you from your belief, I established my own reason for disagreement and advised you of my concerns, and offered you the opportunity to continue with your belief as you saw fit.
So we are all thinking wrong and you are right?
What kind of discussion can you have when your support consists of this? I shouldn't have been required to spell all this out for you...you should have provided indication that you recognized this (if you did) as part of your disagreement, or provide something tangible as support for your stance, since it was your decision to maintain a dispute about it right now rather than wait for further analysis. If you didn't recognize this, it illustrates the problem with having the site do your thinking for you, and I hope you understand better now what my prior posts to you meant in that regard.
Perhaps, but till you can prove me wrong (as you say it needs more investigation but to provide some facts to your assertions), i'll stick to my point: It seems that the PS2/VS2 issues are resolved in the 44.** drivers on the NV35.
I warned you to be careful with interpreting the screenshots to indicate that the NV35 all of a sudden gets higher performance while fully floating point processing, and that integer processing concerns suddenly disappeared. I recognize that you might not care to share that concern, but that you didn't provide any coherent basis for that, so mentioned why I disagreed and tried to avoid a discussion where I simply rehashed the simple observations that progressed us no further towards resolving, but you decided to persist. I don't think it was a useful decision.
PS: Why didn't Nvidia bother to tweaked the tea pot from Ati before? Would have been more easy than the 3DMark03 tweakings.
You, again, did not read the things I rather directly stated before.
To repeat them:
I'm saying that it shows that the NV3x can render at high quality, which is something we knew before.
The teapot screenshot indicates nothing about the performance while doing that properly, it only indicates that the NV35 path in the 44.03 renders properly for applications other than 3dmark03, and an application for which we have no performance metrics at all. This does not tell us anything new.
I further tried to explain the simple observation as to why the GT 4 screen shot still didn't answer that question that I just repeated above. Nor do the absence of PS 2.0 screenshots, when performance results are provide, help.
I don't think any of these details were very difficult to understand, nor do I think any of your comments do anything to address those concerns. Can you tolerate my reasons for finding fault with the screenshots and your conclusions until we have further analysis available? Can you recognize at all that these reasons are not answered by your commentary or the screenshots you've provided so far?
If so, thank you, and please try to give more thought to replying in future instead of just repeating conclusions someone else offers with such incomplete support as you included from that site. I do not require that you agree with me, just that your statements persisting in disagreement with me specifically indicate some understanding of what I said.
If not, please actually include in your post the reasons for why you think so that I have not sucessfully recognized and addressed, to prevent just simply repeating opinions back and forth.
