PDA

View Full Version : Is Current Geforce 6800 ULTRA/GT weak in Shader 3.0?


Pages : [1] 2

sandeep
04-20-05, 02:39 AM
Hi there,

I have heard from some people over forums saying that the current 6800 GPU is pathetic at shader 3.0 particularly in dynamic branching??? Also that these GPUS have some outstanding issues with shader 3.0.

Can someone elaborate on this.

Abba Zabba
04-20-05, 02:47 AM
:(
/nuff said

gstanford
04-20-05, 02:53 AM
ATi and some of their puppet managed media outlets (hello Dave - are you happy Rev has left B3d?) would certainly love you to think that 6x00's SM3.0 performance is weak.

Fact is there is nothing yet that it can be compared to and until there is it is best of class by default.

100fuegos
04-20-05, 03:32 AM
(...)Fact is there is nothing yet that it can be compared to and until there is it is best of class by default.


Amen, that is what I was about to say when I read the title of the thread.

5150 Joker
04-20-05, 04:05 AM
Reverend left B3D? Well only good things can come from that.

gstanford
04-20-05, 04:13 AM
Reverend founded Beyond3D.

NoWayDude
04-20-05, 04:22 AM
At the moment, the NV40,44 and all derivates ARE THE FASTEST SM3.0 HARDWARE on the planet, that joe public can buy
When new hardware appears that can do SM3.0 (and i don't mean vaporware, well wishing, and colective paranoia on forums) and we can see tangible benchmarks, then we can discuss the pros and counts of the NV4X SM3.0 implementation

Fotis
04-20-05, 09:40 AM
They're not weak in Shader 3.0 but they are very weak in HDR which matters most IMHO.

NoWayDude
04-20-05, 09:46 AM
They're not weak in Shader 3.0 but they are very weak in HDR which matters most IMHO.
Previous quote of mine aplies for HDR also.There is no other Hardware doing HDR as we speak.
An i think HDR is not a requirement for SM/PS 3.0, but might be wrong.

Fotis
04-20-05, 10:01 AM
Previous quote of mine aplies for HDR also.There is no other Hardware doing HDR as we speak.
An i think HDR is not a requirement for SM/PS 3.0, but might be wrong.
R300(HL2 demo) can do HDR but not FP16 like NV40.R520 is around the corner anyway.
I never said HDR is shader 3.0 just wanted to say that I consider it more important.

NoWayDude
04-20-05, 10:03 AM
R300(HL2 demo) can do HDR but not FP16 like NV40.R520 is around the corner anyway.
I never said HDR is shader 3.0 just wanted to say that I consider it more important.
NP, took wording out of context
Now HDR with functional AA.... ;)

wEEt
04-20-05, 10:34 AM
Now HDR with functional AA.... ;)
With what card? Not even a R520 can make that possible - maybe it'll support it but it defenitely won't be fast enough.

I guess in a bit more than a month we'll all see how well Nvidia implented SM3 in the NV40.

jbirney
04-20-05, 10:41 AM
Reverend founded Beyond3D.

Incorrect. It was another Dave B (Buron IIRC on the spelling of it) that founded it (and I think Kristos also help set it up) Rev came over after his days on the his 1st Pulpit were winding down....


The Qs of if its weak in PS3.0 is something you can't really answer as its the only thing thats public in PS3.0 In order for it to be "weaK" or "strong" you need to compare it to something. There is nothing to compare it to. So techincally its the fastest SM3.0 part...its also the slowest...

vX
04-20-05, 10:50 AM
Seeing as the unreal 3.0 engine was using 2 6800 Ultra's in SLI mode and still chugging in spots in the demo, (and is a full ps3.0 game) I would assume a sinlge ultra/gt does not have much ps3.0 power to it. I look at ps3.0 on nvidias current cards the same way ps2.0 was on their fx series. :thumbdwn:

The creators of Age of Empires 3, which is a fully ps3.0 game as well, stated the current hardware out now isn't enough to run their game at full settings, so while many will scream FUD at me, because they don;t want to admit their $500 dollar card isn't as " future proof" as they thought, the truth is, nothing is future proof in the world of pc's and nvidia's nv40 series simply has sm3.0 right now just to have it and throw it out there as a huge feature. It does some good though, come next cycle of cards, nvidia will be in it's 2nd generation of sm3.0, so it should have a lead on ati.

Sm3.0 used for speedy shader purposes only can run fine on current tech, but if you're talking running a full blown ps3.0 game, you'll have to settle for 640x480 or 800x600 to get decent frames. This is my guess, as yes, there is nothing we can test it with, but going by word of developers, it seems ps3.0 this time around is just an extra feature for nvidia to toute about on their box.

wEEt
04-20-05, 10:54 AM
Seeing as the unreal 3.0 engine was using 2 6800 Ultra's in SLI mode and still chugging in spots in the demo, (and is a full ps3.0 game) I would assume a sinlge ultra/gt does not have much ps3.0 power to it. I look at ps3.0 on nvidias current cards the same way ps2.0 was on their fx series. :thumbdwn:

Mark Rein:

"For our Presentation we used an Athlon 64 3200+, 1GB of RAM and a GeForce 6800GT and the Demo ran perfectly."

vX
04-20-05, 11:30 AM
i swore he said 2 ultras in sli.....well sh*t, I stand corrected! I hope it does run on one gt, less $$ for an upgrade.

Abba Zabba
04-20-05, 11:45 AM
Of course the GF6800 is the fastest available card when it comes to producing PS3.0 effects since the only other card that supports them is a 3D lab one and is not game oriented.
Having said that, PS3.0 IS SLOW ON CURRENT HARDWARE, face it, the only advantage that I've noticed is when dealing with extra long shaders that use more than 8 textures sample at once (which is super expensive). Now if you consider 50 FPS a gigantic improvement over 40 FPS then I guess you can have you can call it whatever you want.
For example in one my demo, I used PS3.0 to bypass phong illumination for pixels that lay outside of the light range, and guess what? The demo ran twice slower than the PS2.0 path. Branching is expensive on current hardware, but the future generation promises to be a lot faster.
I mean think about it guys, when branching was introduced to the x86 architecture it was slow but then it's gotten much better. I applaud Nvidia for being the world leader when it comes to implementing new features (that's why I'd do anything to get me even an internship with them), however, I don't see too many things I could use PS3.0 for.

MUYA
04-20-05, 11:49 AM
When u have noting to compare it how can u say it's weak? we will see with the next ropund of GPUs. This is null thread IMO and I am tempted to close it

Abba Zabba
04-20-05, 11:57 AM
When u have noting to compare it how can u say it's weak? we will see with the next ropund of GPUs. This is null thread IMO and I am tempted to close it
Hey man, saying PS3.0 is weak doesn't require a comparaison to any other hardware. What I'm saying is, when using conditional branching to avoid heavy per pixel operations, I was deceived by the somehow slow result by comparaison to what regular PS2.0 can do.
Having said that, I already pointed out the advantage of the current PS3.0 implementation on Nvidia hardware when it comes to multiple texture fetching etc...
Bottom line is, PS3.0 could be useful for TRUE softshadows and fractal computation, and that is enough reasons for any game developpers to use it.
Maybe I should put together a demo and post it here to convince you guys of my observation?

Fotis
04-20-05, 01:03 PM
With what card? Not even a R520 can make that possible - maybe it'll support it but it defenitely won't be fast enough.

You seem to know so much about R520!Are you an ATI engineer by chance? :rolleyes: :D

s-flow
04-20-05, 03:02 PM
They're not weak in Shader 3.0 but they are very weak in HDR which matters most IMHO.
:eek: Im impressed with the HDR performance in BIA, SCCT and FarCry..

Clay
04-20-05, 03:09 PM
Hi there,

I have heard from some people over forums saying that the current 6800 GPU is pathetic at shader 3.0 particularly in dynamic branching??? Also that these GPUS have some outstanding issues with shader 3.0.

Can someone elaborate on this.
I did kind of a quasi-comparison regarding dynamic branching with Humus' Dynamic Branching Demo a while back.

Here is a relevant tidbit from Humus that is needs to be read before reading the results below:
Notice that the default settings are different between the ATI and NV cards. I simply choose the path that's fastest on each card. For ATI cards it's faster to just zero stencil when the stencil test passes, while on NV cards this seems to interfere with early stencil rejection, thus they are faster when I clear the stencil buffer between lights instead. That's still quite fast on ATI, but not as fast as zeroing.

http://www.nvnews.net/reviews/bfg_geforce_6800_ultra_oc/page4.shtml (near the bottom of the page)

http://www.nvnews.net/reviews/bfg_geforce_6800_ultra_oc/images/humus_dynbrnch.gif

This was my conclusion at the time:
The results appear to fall in line with what Humus stated above. Naturally, a card would be directed to operate at the optimal settings given that no image quality degradation would occur. There were no IQ differences at all that I could detect. We're talking about a stencil test here though and not textures, AA, AF, etc...so it's not a surprise that IQ remains constant. What I found to be most interesting is that, given each card's best approach, the BFG 6800 Ultra OC was 10%~15% faster than the X800XT PE. The worst-to-best comparisons are quite different but they're also quite irrelevant.

wEEt
04-20-05, 03:10 PM
You seem to know so much about R520!Are you an ATI engineer by chance? :rolleyes: :D
I will give you 10$ if you can play any modern game @ >1024x768 w/ 4xAA and 8:1 with HDR activated. I don't even think that this will be possible...

Acid Rain
04-20-05, 03:59 PM
Having said that, PS3.0 IS SLOW ON CURRENT HARDWARE, face it, the only advantage that I've noticed is when dealing with extra long shaders that use more than 8 textures sample at once (which is super expensive). Now if you consider 50 FPS a gigantic improvement over 40 FPS then I guess you can have you can call it whatever you want.
How people come to the conclusion that PS3.0 is slow is beyond me. It's already been established that performance is higher when using 3.0 to achieve the same or similar output result that 2.0 or 2.0b is capable of. It has been seen that you can use 3.0 as a performance enhancement. How does this make 3.0 slow?

I've seen it claimed that "3.0 is soo slow" many times now, usually by people who haven't even run a 3.0 card :rolleyes: , and that claim is simply unfounded. Sure, newer hardware may realize even greater performance than current tech using 3.0, well, that's a given, but it's already a better way to do things even with current NV 40 tech. Again, how do people come to the conclusion that it's slow??? Why does my Far Cry run better with it??? Should I be forcing 2.0 and call it a day? Hardly. I like the INCREASED performance too much.

Now, correct me if I'm wrong, but after your claim that 3.0 is currently slow, it appears that you've turned around and contradicted yourself with the 40 fps vs 50 fps statement. If it's slower, why would the 40 vs 50 fps comparison even be made, unless you are saying that 2.0 is the doing 50 and 3.0 is doing the 40? Also, that's quite a reasonable increase, don't you think?

Abba Zabba
04-20-05, 04:37 PM
Please reread my posts as I've never claimed the current implementation of PS3.0 is always slower than the 2.0 counter part.
I've specifically said that in certain situations when we're dealing with multiple TEXTURES FETCH like in the soft shadows generation, PS3.0 comes out victorious because it saves the card the trouble of seeking tremendous amounts of texture data from the cache.
Here's a more in depth explanation;
In the most straight forward soft shadows generation, we fetch (retrieve) multiple depth samples from the caches, say sample0-N (N is 64 in my case) and for every sample k (where 0 <= k <= n) that is greater than the stored distance (or depth) value, we substract 1.0/N from the current illumination value

Let's say that from 0 to k = 8 all the samples indicated that the current fragment is NOT occluded, then we decided it is useless to fetch the 64 - 8 = 56 other samples, declare illumination 1.0 and move on to the rest of the code.
With PS2.0, there is NO conditional branching and hence once we got on the track to retrieve the first 8 samples, there is no way to break away from it which results in a whooping 64 texture fetch ops.

Now with a little background on how memory buses and cache work, you'd figure out how COSTLY and expensive a texture fetch is, that outrageous expense is very often translated by low FPS.

With relatively simpler shaders, like Phong illumination, we might introduce branching to avoid applying the Phong equation (one of the most expensive things as it calls for vector normalization and a reflection) to the fragments laying outside the range of the light source. According TO MY EXPERIMENTS, using PS3.0 in this situation results in LOWER framerates than the PS2.0 path on a GF 6800.

Having said that, the only reason why I got myself a GF6800 is the VS3.0 and PS3.0 features thatthat none else other than 3D labs had.