supposed nv30 shots. where's the journalistic integrity? [Archive]

array

29-Jul-2002, 08:05

ok, honestly i think this situation is ludicrouse. ben6 claimed that this was rendered in real time on an nv30:

http://www.exluna.com/products/gallery/killeroo.jpg

i dont care WHAT nvidia told him. its not possible, and wont be for a LONG time. that kangaroo image was rendered using a path tracing algorithm. such renderings can take up to days on modern processors. we'll have real time photon maps before we have real time path tracing (due to the slow and noisy nature of the path tracing algorithm).

you can find more info on what path tracing actually does here:

http://graphics.stanford.edu/~henrik/papers/ewr8.html

more info on photon maps here:

http://graphics.stanford.edu/~henrik/papers/ewr7/ewr7.html

it would be amazing (but possible) if the nv30 could do realtime ray tracing. path tracing is still a LONG way off. the only possible way that kangaroo could be rendered in real time is if nvidia utilized a baked texture in their demo to fake global illumination.

DemoCoder

29-Jul-2002, 08:40

There's no reason a similar image can't be generated by cheaper methods. NVidia can't run the (full) Final Fantasy the Movie in real time, but you can do pretty good cheats, and the art of real time graphics is cheating, not correctness. That didn't stop NVidia from including the REAL SquareSoft images in their presentations.

Most people won't be able to tell the difference between path tracing, radiosity, ray tracing, photo mapping, or scanline rendering. If Nvidia shows a KillerRoo that is done via scanline tricks and hand-coded shaders and lightmaps, is it a lie to say point to that image and say they generate it in real time?

ATI is demoing "Lord of the Rings in real time". Does that mean ATI is taking the art assets from WETA and directly compiling them unchanged to render on the 9700. Their terminology sort of implied that they are rendering the exact same scene and artwork.

We know this is bullshit, since LOTR was rendered on a huge renderfarm, using gigabytes of geometry and textures. But I have no doubt that they have a cut down and hacked real time demo that looks close enough.

I'll bet that there will be NV30 demos that demonstrate all of the ExLuna content winners in real time. They won't neccessarily be using ray tracing, radiosity, or path tracing, and the other goodies that Exluna/BMRT have, but with appropriate tweaking, they will look close enough for the amateurs.

It's possible someone at NVidia sent Ben a link saying "NV30 will be able to render images like the one at this link" Which is in fact possible. NVidia did not say to Ben "NV30 will be able to do path tracing in real time"

array

29-Jul-2002, 08:50

good point. but LOTR and FinalFantasy the movie were rendered in PRMan, which is a scanline renderer. the fact that modern day cards cant render those scenes in real time just reinforces how far off we really are in terms of rendering scenes like the killeroo. ben6 should at least clarify that those images were faked before sites like nvnews gobble up his words and post them on the front page.

Randell

29-Jul-2002, 08:52

I know its been said before, but ATI have done the exact same thing with their crystal ball & animusic demo's.

Both major IHV's are claiming this kind of content is feasible on their DX9 parts.

end of surely.

array

29-Jul-2002, 08:54

ati claimed "real time" ray tracing and HDRI with their demos. both are a far cry from the cpu power necessary to computer path tracing.

LeStoffer

29-Jul-2002, 09:10

We still don't know whether nVidia claimed that the pics was from an actual NV30 board (running in real time) or whether it was just a software simulation based on the NV30. It could also be something along the lines of "this is want we know it will be able to do, because we did a couple of simulations and the math is there".

Ben6 said that he have asked nVidia for clarification. Maybe we should wait for the info before jumping to conclusions?

Dio

29-Jul-2002, 09:58

Fast, even real time ray tracing is possible on R300 - it was discussed in some detail at Siggraph. One presumes it is also perfectly possible on nv30.

array

29-Jul-2002, 10:09

Sharkfood

29-Jul-2002, 10:13

ATI is demoing "Lord of the Rings in real time". Does that mean ATI is taking the art assets from WETA and directly compiling them unchanged to render on the 9700. Their terminology sort of implied that they are rendering the exact same scene and artwork.

The question is, what imagery are they "demoing"?? The real output of the quoted technology, or LOTR?

It simply can't be compared of two situations that basically fall down to:
1) Company X produces Y, that is supposed to look like Z (but isnt exactly) and showing Z.
vs
2) Company A produces B, that is supposed to look similar to C (but isnt exactly) and showing B.

I'm unsure if ATI showed/propagated images of the real process or of actual LOTR, but it's obvious NVIDIA is showing/propogating imagery that is NOT the target of their output results. This is deceiving and misleading.

artisan7

29-Jul-2002, 10:18

Most people won't be able to tell the difference between path tracing, radiosity, ray tracing, photo mapping, or scanline rendering.
"

and COmposting too.
all movies like Final fantasy/Toystory/Lord of the RIngs and many others
movies done with Pixar's renderers or others projects done with GI /mental ray ,use heavy COmposting work
to SAVE time or to improve the details at some scenes.
..in other words is a "FAKE" of the photorealistic effects with "Realistic Lighning" and "realistic" shadows".

Nv30 with all that precision 64bits/128bits plus all that programable pixels/vertex shader should not have any problem displaying
"offline rendering" any shot from any movie... without the need to
do composting works.
however real time animation in more than 30Fps+ with
that kind of Quality is a diferent Story .. something i like to see someday.. :)

what it is interesting is .

1)that Nvidia CEO has told many times diferent interviews "we are going to make obsolete the cpu" ... huh?
2)Nvidia bought ExLuna ,main competition to Rendermans favorite Holywood renderer.
3) the very High precision in colors ,
plus the Highly programable GPU of Nv30... and incoming products.
4)Plus the really good comments from JC about Nvidia "Far sight in the future" making possible "offline renderings" in Nv30.
5)Nvidia shading language Cg...
6)Nvidia demostration of a lower quality final fantasy movie in a geforce4?

Mmm.. make me wonder if there is
a WAY or posibility that "software(cpu) rendering" in a couple of year
could be really Obsolete! like Nvidia is claiming .
Now that they owns exluna main competition for renderman
that should not be impossible ,very difficult maybe ,but not totally impossible ,remember that what we see in Holywood CG movies ,
use lots of effects "fakes" too.

That will really be Incredible!!! hollywood quality at home ,
it will not be real Global Ilumination or radiosity or Mentalray realistic lights photons ,it will be still "fake" the offline rendering made by the NV30
Nv40's.. but the final result can look ,AS GOOD as the final shot done
for CG Holywood movie.

any Computer Graphic studio will not think TWICE to switch to offline
rendering (whenever its possible) while working in a movie ..
if the final quality of the scene is the SAme... wow! time will tell ..
but this possibility it something i will like to see someday!

array

29-Jul-2002, 10:28

just to clarify some things regarding renderman:

1) Renderman is a standard.
2) PRMan is Pixar's implementation of the Renderman standard.
3) A man named Larry Gritz made a FREE implementation of the renderman standard while he was in college. He called his renderer BMRT.
4) Larry Gritz gets a job at pixar.
5) Larry Gritz leaves pixar and helps form Exluna and launches flagship product, Entropy.
6) Steve Jobs (owner of pixar and lord of all things evil), being the prick he is, sues Gritz and his company over violating trade secrets. Supposedly Gritz stole an algorithm (stochastic sampling) which just happens to be widely available on the internet :roll:
7) Nvidia buys exluna and settles the suit with Pixar. Part of the settlement was that BMRT be discontinued (the download link on www.exluna.com is gone), and that development of entropy be frozen.
8 ) You're reading this post and wondering what happens next :)

i just thought it was relevant cus we're talking about Renderman and not a lof of people are up to speed yet :D

KnightOfDarkness

29-Jul-2002, 10:30

array

29-Jul-2002, 10:40

Okay, this is kinda not related to the exact topic, but lemme get this straight.

So all GPU's and game systems now are scanline renderers right? And then the next step up from scan line rendering is what? Photon mapping or ray-tracing? Or is the former just an advanced form of the latter? Then after ray-tracing, the next step up is radiosity right? This is a technique that includes indirect lighting, if I am correct, meaning light that bounces off objects and then illuminates other objects is also traced when the scene is rendered? And from this thread, should I assume that the most advanced, time-consuming and accurate form of rendering we have at this age is path-tracing? Is there a method even more accurate and more strenuous to implement? Ne way, it'd be nice if someone could clear this up, as I am sure some others might benefit from the reply besides myself :p.

Ray Tracing is an algorithm which traces rays from the camera (viewer) to geometry, then calculates pixel color based on the normal of the polygon that the ray intersects and it's relation to a lightsource.

Radiosity is an algorithm which subdivides a mesh into patches and then calculates radiance. it is sparcely used due to it's reliance on mesh topology (a dirty mesh will make a dirty image) and the fact that rendering times are directly proportional to mesh density within the scene. it can only simulate the most basic global illumination effects, such as color bleeding and soft shadows.

Path tracing is an extension of ray tracing which traces rays from the light to the camera. it is VERY noisy and VERY time consuming. it can simulate some cool things like caustics and global illumination though.

Photon mapping is the latest and greatest thing. it generates images by casting photons from lights, and then generating a geometry independant(BIG plus) photon map (hence the name). photon mapping is rediculously fast compared to path tracing, and is virtually noise free when properly implemented. as you may have guessed, photon mapping is based on physical behavior of light, hence it can simulate every natural phenomena EXCEPT: diffraction, interferance and polarization. hope this helped.

if you want more info, pick up a copy of this book:

http://www.amazon.com/exec/obidos/ASIN/1568811470/qid=1027935530/sr=8-4/ref=sr_8_4/103-2305496-1881415

or read the papers on this site:

http://graphics.stanford.edu/~henrik/

if you want examples of images rendered with photon maps check out these sites:

www.finalrender.com
www.splutterfish.com (the beta version of this renderer uses path tracing, but the commercial and final versions use photon mapping)
www.vrayrender.com
www.mentalimages.com

hope this helped :D

Dave B(TotalVR)

29-Jul-2002, 11:40

Ilfirin

29-Jul-2002, 12:27

Ray-tracing on Programmable GPUs (http://graphics.stanford.edu/papers/rtongfx/)

Yes, that article is posted all too often for these types of debates, but it is still relevant ;o

jasonjlee

29-Jul-2002, 12:29

array

Brilliant clear post full of info. Thanks.

[edited for typo. Blimey I think I need another coke]

Qroach

29-Jul-2002, 12:58

Array, have you read through that article Ilfirin posted before? I'd like to hear your comments after reading it...

eSa

29-Jul-2002, 12:59

I don't know why some of the people here get so worked up about Ben's posts. There is no magic here, nVidia is simply leaking some non-realtime pics rendered using nv30. If the "killeroo" is indeed claimed to be real-time, well you can always fake things... I would say that with very high quality HDR enviroment map + some per pixel stuff you could render the "killeroo" close enough to the original path traced quality. Actually, you could probably build HDR envmap directly from original global illumination solution.

Someone mentioned the ATI's "crystal balls" demo in post above. There is Siggraph 2002 whitepaper from ATI which documents the techniques used to get similar quality than the original 98' Siggraph HDR animation. No magic there either, just straighforward code.

Whole point here of course is that people "less expirienced with gfx algorithms" :wink: could believe that nv30 will deliver games with similar quality... That hype may just backfire and nVIdia could lose some credibility.

sancheuz

29-Jul-2002, 13:18

Where is Ben during all this? Those images by the way are not similars, they are the exact same image posted on the exluna site.

eSa

29-Jul-2002, 13:19

Ah, sorry could not resist :D

This image (from finalRender pages) is done with HDR enviroment map as only source of illumination :

http://www.finalrender.com/finalrender/images/helm8.jpg

Quality isn't bad... And you could pull of that in realtime with R300/nv30.

Allthought there still some time to go until we get here (in real-time that is) :

http://www.finalrender.com/finalrender/images/fr_glass_ren.jpg

nggalai

29-Jul-2002, 13:46

Hi eSa,
I don't know why some of the people here get so worked up about Ben's posts. There is no magic here, nVidia is simply leaking some non-realtime pics rendered using nv30. If the "killeroo" is indeed claimed to be real-time, well you can always fake things...
That's an important point some may have got out of sight. Neither NV nor Ben's contact claimed that said images were rendered in "real-time," but previous information strongly indicates that NV is aiming also at the CGI market, providing NV3x as an off-line HW render solution reminiscend of ART's RenderDrive.

ta,
-.rb

Mulciber

29-Jul-2002, 13:51

rubank

29-Jul-2002, 14:19

(looking at eSa:s link..)
Hmmm,
the wine bottle is less than half full, yet there´s not a simple trace of wine in the glass. Is this guy drinking from the bottle?
Nah, it must be a fake :D

Dio

29-Jul-2002, 14:36

dio- i said RAY tracing was possible in hardware in my original post. it's PATH tracing which we wont see for some time, or ever for that matter since photon mapping can do everything path tracing can
more efficiently and with less noise (i.e. artifacts)

Sorry; I never implied that I was talking about path tracing, and am aware it is much more difficult. I was answering other people who were discussing if ray tracing was possible.

Dio

29-Jul-2002, 14:53

DIO:

"Fast, even real time ray tracing is possible on R300 - it was discussed in some detail at Siggraph. One presumes it is also perfectly possible on nv30."

The tracing of a ray is most likely possible, real time ray traced scenes, youve gotta be kiddin. For each ray you need to re-transform the whole of the geometry millions and millions of rays are traced per scene.

I do not believe this is correct. Fast ray tracers create a geometry database to avoid comparing with more than the smallest possible portion of it.

I report only what I have heard from people who were at Siggraph, who said that the ray tracer running on what I believe was an R300 was very fast and very impressive.

I refer once more to the Purcell et.al. paper included above... it was his demo.

Gollum

29-Jul-2002, 17:11

This thread is interesting and insightfull, which makes it all the more weird that it all started off with a false accusation:

ok, honestly i think this situation is ludicrouse. ben6 claimed that this was rendered in real time on an nv30:
http://www.exluna.com/products/gallery/killeroo.jpg
Nobody else willing to defend Ben or am I the only one who got curious about when he supposedly made that claim? Turns out its simply not true, here's what ben said:
I was skeptical too , until I saw the kangaroo in one of those screens on that page in motion with refraction and irradiance .
there's videos included in the package , one of aging wood, one of a kangaroo, both of which are on Exluna's page except not as videos but as a still render .
http://www.exluna.com/products/gallery/killeroo.jpg
The difference being that it was fully colored and textured in a room with a Nvidia sign in it, instead of the simple model in the screenshot

It doesn't take a rocket scientist to understand that what ben said was not a claim of NV30 rendering said image, but rather that the killeroo shown in that image was textured, shaded and animated in a video he got sent by Nvidia (which by description sounds like a typical real-time demo to me, so it might even have been recorded off an early NV30 board). Basically it appears they used the geometry of the killeroo for a different scene. Quite an important difference! I suggest people should read more c a r e f u l l y before making accustaions IMO...

To sum it up, neither Nvidia nor Ben claimed that specific image was rendered on NV30! As much as I apreciate the discussion about various rendering algorithms (highly interesting to me as a 3D artist) and which of them might be feasible in realime and which not, the base of the thread kinda spoils it as it gives the whole thing a utterly unneccessary negative touch.

array, you appear to know a lot about rendering and I greatly enjoyed the info in some of your posts in this thread, why not use that knowledge in a direct discussion about the issues surrounding real-time raytracing or other redering slgorithms, instead of making false accusations first?

Nite_Hawk

29-Jul-2002, 18:30

Gollum:

I partially agree with you, especially in that Ben doesn't need to inform us of anything, and it's great that he provides us with various tidbits of information and rumors and such.

On the other hand though, take a look at the title of the thread. "Early screenshots taken on NV30", now there seems to be a lot of question as to wether or not those images actually were taken on NV30 as they appear to be scaled versions of the same ones provided by exluna. (created with Maya and such). Some of the attitudes here are really negative and that's a bad thing, but the questions raised about the screenshots are valid imho. Can the NV30 actually produce those exact images, or will they be aproximations of those images?

With the Killaro, the post was somewhat confusing, as Ben didn't mention that the image provided was not produced with NV30, but only that there was a video which seems to have been and differed in that "...it was fully colored and textured in a room with a nvidia sign on it". Having read the post, I agree that Ben never said it was rendered on NV30, but I could see how people might mistake that it was, especially given the title of the thread. Still, I don't think that Ben meant to fool anyone, only that he should have taken a bit more time to explain where the images came from.

Array:

Sounds like you've done a lot of rendering work! :) It's always nice to find someone who's up to date on the scene. (I've been out of it for about 2 years now, haven't had as much time since I met my fiance!) I did a lot of work with Radiance for a couple of years, and povray before that. Read the papers by Henrik a while back too on Photon Mapping; it's good stuff.

Do you happen to know if it suffers from having large and small objects in the same scene in the same way that stochastic raytracing does? Basically when using radiance, if you have a really big scene (in my case a cathedral) and if you want to have small details (like say the face of a statue), you need to set your ambient resolution up *really* high because otherwise the density of rays that get sent out is too low, and small objects end up with inconsistent shadows and lighting. I remember when I was playing with photon mapping in povray that it has larger and smaller photons, seemingly for close up and farther away objects. Do you happen to know if this is the case, and if it fixes this problem?

Nite_Hawk

Pavlos

29-Jul-2002, 20:56

Actually, you can accelerate global illumination with upcoming hardware (R300-NV30). If you have a method to accelerate the ray database traversal and the intersection tests, like the method described in the Purcell’s paper, you can accelerate any Monte Carlo algorithm (like Path Tracing).

As for the actual images, they can be produced on any DX9 hardware, but keep in mind that the translation of renderman shaders to ps2.0 shaders is not trivial and requires a lot of engineering effort (they are not so similar that most people thing). I doubt Exluna can produce in a few weeks (or even months) a renderer that performs this translation.

Edit: I’m referring to the Exluna pics, not this one. This one is much easier to reproduce, even with real path tracing.

Just to clarify some things about Global Illumination:

Global Illumination algorithms (or better Light transport algorithms) can be roughly divided into two groups: Monte Carlo methods and finite element methods.

The Radiosity algorithm belongs to Finite Element methods. These algorithms, as mentioned by array, are based on dicing the scene into patches. They are extremely fast compared to MC algorithms, but they solve only part of the problem (only diffuse environments), they need a lot of memory, they work only with polygons etc…

Path Tracing, Bi-directional Path Tracing and the Metropolis light transport are some of the most important Monte Carlo algorithms. MC algorithms are based on Stochastic Sampling. They are robust, they use very small amounts of memory, they are scene independent etc… but they are extremely slow (if you don’t like noise you have to wait). The MC algorithms mentioned before are called unbiased because the only error you can find in the final image is noise. Noise denotes insufficient sampling and you can remove it by using more samples. The great challenge is to get less noise with a given amount of samples. Metropolis light transport is the best method we have so far.

There are also the biased MC algorithms, like Jensen’s Photon maps and Ward’s Irradiance caching, which try to speed up the calculations by caching and reusing results. The final image is noiseless but caching introduces error (bias) and you need infinity samples to get zero bias. But things look good with just a few hundreds of samples. Commercial renderers (like BMRT, Brazil, Entropy, Arnold and others) use variations of the last two algorithms (usually Irradiance caching for diffuse reflection (diffuse to diffuse light transport) and Photon maps for caustics (specular to diffuse light transport)).

Ps. Note that English is not my native language. Also I have oversimplified things and skipped some categories to keep the message “short”.

nggalai

29-Jul-2002, 21:05

Hi Pavlos,

<snip>

Ps. Note that English is not my native language. Also I have oversimplified things and skipped some categories to keep the message “short”.
Still a great read. :) Thanks a lot.

ta,
-Sascha.rb

array

29-Jul-2002, 21:07

pavlos- nice post :) but afaik, Entropy only employed path tracing to computer GI and caustics. if you want photons in a renderman compliant renderer, give AIR a try, or wait for PRMan11.

Oompa Loompa

30-Jul-2002, 00:34

Photon mapping is the latest and greatest thing. it generates images by casting photons from lights, and then generating a geometry independant(BIG plus) photon map (hence the name). photon mapping is rediculously fast compared to path tracing, and is virtually noise free when properly implemented. as you may have guessed, photon mapping is based on physical behavior of light, hence it can simulate every natural phenomena EXCEPT: diffraction, interferance and polarization.
Hmm. Do you mean that photon mapping only simulates the particle properties of light, and that physical phenomena dependent on its wave-like characteristics can't be simulated by photon mapping?

Gollum

30-Jul-2002, 00:41

mmmh, this thread sure turned out having a good number of very insightfull posts, one of the best here for days IMO... :)

Nite_Hawk, I really do agree with you, it is abgsolutely legitimate to question the authenticity of the real-time nature images and wether they were really rendered on NV30 hardware, on a simulator or neither. Nvidia sure has caused some confusion here and needs to clear things up or it will cast a negative light upon them and NV30. I just felt the need to express my dissapointment that this thread had to start out the way it has, by attacking ben for posting Nvidia's images. If anybody, then nvidia shoudl have been the target. Sure, as journalist Ben could have chosen not to post the images before he had 100% confirmation that these were rendered on NV30 hardware, but since they apparently came directly from Nvidia I can understand his decision to post them. Hope this issue will get cleared up in the near future.

Pavlos, excellent post there, gotta read it a couple more times to understand it all though. Do you happen to know what (possibly mix off) algorithms LightWave 3D's radiosity employs? I can't seem to find any detailed information about it...

cellarboy

30-Jul-2002, 01:06

Do you happen to know what (possibly mix off) algorithms LightWave 3D's radiosity employs? I can't seem to find any detailed information about it...

Lightwave uses the Monte Carlo Radiosity Algorithm. I'm not sure if this is exactly the same thing, but there is some info here:

http://www.cs.kuleuven.ac.be/cwis/research/graphics/CGRG.PUBLICATIONS/HMC/hmc.html

Gollum

30-Jul-2002, 02:02

Thanks cellarboy! :)

30-Jul-2002, 04:00

LeStoffer

30-Jul-2002, 06:58

One of the best methods for real-time rendering using global illumination is spherical harmonic environment maps. The killeroo shown could easily have been rendered in real-time using this method on next generation DX9 hardware (even some current generation hardware). For a good paper on SH environment maps see:

http://research.microsoft.com/~ppsloan/shillum_final23.pdf

Very interesting stuff there, thanks SA!

I happen to think that soft shadows are crucial if we're going to have near-cinematic quality, but I didn't know that we were so close. http://www.plauder-smilies.com/biggthumpup.gif

SA, what kind of performance advantage will DX9-level hardware have
over the required PS 1.4 (Radeon 8500) and simple dot product instructions? Can the transfer of matrix transforms and SH-rotation be performed on a DX9 GPU, so we will have a interactive real-time rendering. Or do we still need some precomputered cube map with DX9?

I guess you could use SH environment maps on a per-object basis so performance won't drop to far (just like you can apply ray-tracing per object in, say, 3dmax).

Edit: More specific question after reading the paper more carefully. :oops: