GeForce FX: 8x1 or 4x2?

I'd also like to add that this mess could have just as well been avoided if nVidia was just upfront about the whole thing from the beginning.

Advertise it as 4 "Uber Pipelines!" or some other such marketing non-sense, and have a nice explanation about how these are "better than your average pipeline!" Talk about how "Doom" really benefits from these pipelines compared to the "normal" pipelines, etc. We';d all be talking about the BENEFITS of this architecture compared to the GeForce4, etc.

We'd even be talking about how SMART it is, to get some reasonable performance benefit without needing the "cost" of the bandwidth needed to fill 8 "complete" pixels per clock.

Instead, you get the results we have here....everyone talking about how this is NOT as GOOD as a "complete" 8 pixel per clock engine. Shame really....but it's nVidia's fault for advertising as 8 pixels per clock....and letting the web p/review sites run with it.

 

yes indeed. thanks Joe. that was the best explaination ive read yet. even I can understand that

 

NV also can improve in long terms. i guess that nv35 will have 2х fp performance and nv40 - 4x over nv30.

its simply question of the priority

.13 and flwxible pipes and then more fp power (or more pipes)
8 power but more simply pipes at .15 and then .13 and then more flexible pipes

but second case it 3 stages while first is two

 

Why cant ATI do this:

more pipes at .15 and then .13 and more flexible pipes?
Besides, its not stages that matter but actual time.

 

One would hope so, given the long production cycle.

 

I call it 4x2+8z

 

yep time mathers you right
but software adoption also valuable part of this time

seems that R350/RV350 got more flexible pipes.

all as i guess before

so we will see OGL 2.0 at 9800/9600

 

Well, I'd like to agree with Steve Martin from an old SNL.....

 

Most interesting...care to elaborate? (Or I guess we could wait until Tuesday....nah!)

I do distinctly recall Dave (I think) dropping some blatant "hints" about programmability in at least the new R350...something along the lines of nVidia not really having anything on ATI in that dept. I guess we'll find out soon enough.

 

the thing is that not everyone can be expected to be "in their right mind," especially when, no offence to Croteam, they built an engine that is not getting licensed much (at all?). not to mention just the oddness of them commenting on the issue at all and the fact that if you are going to discount the value of 8x1 you might as well downplay clockspeed, memory bandwidth, and all the other relevant numbers as well. on a side note, i imagine there was a lot of people said it was immaturely ridiculous to assume J. Edger Hover dressed in drag as well.

*slightly edited for clarity.

 

Hmmm...did someone say more programmability?

 

Isn't that exactly what we should do ? (downplay clockspeed, memory bandwidth). Think about the Kyro 2 f.e.

And, they're not discounting it, just saying that there are something that are much more important and that's actual performance/features.

 

well John Reynolds showed me a pdf on the kryo2 a while back and from what i got from that; it sucks.


seriously though, performance/features are all that matters; but the simple fact but its not like there are little invisible elves that run around and make things happen; there are real reasons for these things. the truth of the matter is that while things like tile based rendering can make up for things like clock speed and memory bandwidth, they are all factors in the equation. i don't want to hear that a lower clock speed and lower memory bandwidth are unimportant unless you can present something like tile based rendering to fill in the gap. so Croteam claims that it doesn't matter that the fx is not 8x1; but until they say what they think is going to make up for that their assentation holds no real value. furthermore, the fact that the cards are currently shipping, but have only been previewed on a few sights and some of those previews are blatantly questionable, does a lot to make one question the validity of Croteam's opinion on this matter.

 

kyleb, yes, of course, you're entitled to your opinion and your justification for such an opinion. Just like anyone else is entitled to theirs'.

 

It really doesn't matter much how many pixel pipelines GeForce FX has. All it matters is how fast can it execute shaders.
Radeon 9700 has 8 pipelines with 1 TMU each. It will typically output 8 pixels every clock if you use pixel shader with only one arithmetic instruction (8 pixels/clock on average). If you have pixel shader with two instructions Radeon 9700 will output 8 pixels every 2 clocks (4 pixels/clock on average) and my test proves this. If we have a pixel shader which is 8 instructions long then Radeon 9700 will output 8 pixels every 8 clocks (1 pixel/clock on average).
You could even have a part with only one pipeline and a pool of 8 TMUs and 8 ALUs that would perform on pair with Radeon 9700 in many situations. Shaders are getting longer and longer and no one is going to stop this. So it doesn't matter if you run a 8 instruction pixel shader and one card outputs 8 pixels every 8 clocks and other card outputs 4 pixels every 4 clocks or even if one card outputs 1 pixel every cock, since they all output one pixel per clock on average.
The problem with FX as can be seen in my test program is that it doesn't seam to have 8 float ALUs. It doesn't even seam to have 8 int ALUs, but only 4 int ALUs and "2.5" float ALUs. Which could be because of drivers (it eventually should be faster per clock then GeForce right?).

 

FWIW... here's an interesting observation by the lead programmer of Splinter Cell (don't ask me for more info coz he and I can't tell) :

 

I hate doing this, but anyway...
http://www.theinquirer.net/?article=8067

Yes, it's The Inquirer. Yes, it's probably BS.
But it would explain a lot if it was true.

For example:
Some MIPS tests showed the R300 could do 1.3 Instructions/pipe/clock - The Inquirer explains that by saying it can do 2 instructions/pipe/clock when one instruction uses 3 components and another 1 component ( scalar )
So, sometimes, the R300 could be significantly faster ( maybe even more than 1.3 ) than a traditional 8x1, and in worst cases it would be on par.

Also, this would explain the excellent integer performance. And the bad FP performance.

It does seem like a rather strange explanation, however. But I'd still consider it as possible.


Uttar

P.S. : After further thinking, the following could be supposed from that Inquirer article: The NV30 got 1 ALU per pipe. That this ALU can actually do *two* operations per clock at FP16, and one operation per clock at FP32. But using TMUs can't happen at the same time as using the ALU.
So, the GFFX should be very fast when using textures + integer. But it should also, theorically, be very fast at FP16 and no textures. And from the little we seen, it isn't. Strange...

 

This is insane, none of you actually know the NV30's architecture 100%. (unless you helped design it)
Everyone has their own theory, what's worse is that nVidia isn't releasing any more info for the time being.

This seems more like a game of pong.

 

That just sounds like the comments others have made in these forums (I think in this very thread)...

With no source mentioned, I don't know if that is independent confirmation of what people have said here, or John Smith is just a (lot) smarter about interpreting info ripped from across the internet than some of the other Inquirer staff. It could even be a developer who conducted such testing contatct John Smith directly, and John Smith is just smart enough to quote them (if you call that "quoting"), since the use of "we" reads like a team of programmers talking.

I suppose there is also the possibility that John Smith conducted the investigation himself, but then I have to wonder what the heck he is doing at the Inquirer...do they make good money there?

 

Well if they'd operate like ATI, I'd provide info that I measured myself on a GFFX instead of theories...