Real time Pixar graphics

[jvd]

As seen here http://www.beyond3d.com/forum/viewtopic.php?t=7991


I find this very interesting from someone in the busniess .

Daliden wrote:
A while back there was some talk about how it will soon be possible to render Pixar-equivalent graphics in real-time.


For well over three years now there's been talk.

Daliden wrote:
There even was a demonstration of real-time rendering, albeit at a low resolution and no AA.


Also albeit with baked, not procedural, textures. Albeit with simplified geometry. Albeit with no motion blur. You are talking about the Luxo, Jr. demonstration at MacWorld in 2001?

Daliden wrote:
Time has passed, but has there been any progress in this field?


Yeah, at Siggraph 2003 I showed a toy ball procedurally shaded in real time (>60 fps) with one (not three) lights. On the plus side, it *was* motion blurred (by an ad-hoc procedural technique, which was simplified for a spinning ball, not a squashing/stretching/bouncing ball. But the method can be generalized to the full solution.) And the geometry, while simplified, was an accurate sphere to subpixel precision.

Also shown at Siggraph 2003 was a *single* uberlight in real time.


So, Reality Check: Toy Story 2 (and Toy Story before it) was rendered on the order of 1/1,00,000 real-time. (See Tom Duff's posting to comp.graphics.rendering.renderman, http://groups.google.com/groups?hl=en&lr=&ie=UTF-8&oe=UTF-8&selm=3909BD4B.A107CD05@pixar.com ) Just let Moore's Law (no imagined cubes required) work and a uniprocessor PC will manage it in real-time 30 years later. Toy Story was released in 1995, so software rendering of Toy Story in real-time on a single processor PC should be possible around 2025.

Assume hardware rendering is on the order of 1,000 times faster than software rendering. That puts be real-time "Toy Story" 15 years after its release, or 2010, still seven years from now.

Reality Check Squared: Luxo, Jr. was rendered 1986. Add 15 years = 2001. Oops, nobody could do Luxo, Jr in real time in 2001, and still nobody has done so in 2003.


There isn't any question it will happen someday. Real-time is getting damn good, but it's still got some to go. But by the time we get there, we'll still have more to go to catch up with today's films, let alone with tomorrow's.

In the meantime, we keep dreaming.

-mr. bill

 

[notAFanB]

Just let Moore's Law (no imagined cubes required) work and a uniprocessor PC will manage it in real-time 30 years later. Toy Story was released in 1995, so software rendering of Toy Story in real-time on a single processor PC should be possible around 2025.

not being very specific here are they?

 

[Paul]

Just because you can't render Toy Story in it's exact complexity on upcoming PC hardware doesn't mean you cannot achieve it's look.

I do think this will be the case in a few years, you'll see similar to toy story graphics; but without the huge rendering load Pixar did with the movie.

 

[jvd]

Just because you can't render Toy Story in it's exact complexity on upcoming PC hardware doesn't mean you cannot achieve it's look.

I do think this will be the case in a few years, you'll see similar to toy story graphics; but without the huge rendering load Pixar did with the movie.

The point isn't if you can do it with all these hacks in it. HE is talking strait quality port of the movie. There is a render that was done in 86 and it is still not being done in real time. I was thinking toy story 1 graphics in another 3 or 4 years but now i'm left wondering if that will be enough.

 

[zidane1strife]

...

The animators who created Toy Story used high end Silicon Graphics workstations to prepare the animations. More than 400 3D models and 2000 texture maps were used, and the two main characters (Woody and Buzz) each had over 700 animation controls, including 212 on Woody's face and 58 on Woody's mouth alone. The modeling and animation preparation took over ten man years to complete. No motion capture was used in the entire film - everything was animated by hand. The final frames were then rendered by a "rendering farm" of 117 multiprocessor Sun SPARCstation 20's. It took 800,000 machine hours to render the 114,200 frames of the 79 minute film.

... and so...
800/114.2= 7.0Hrs... we need to do things 757K times faster to do that at 30fps OH no!!!! but wait, wait what did they have back then?

So, Reality Check: Toy Story 2 (and Toy Story before it) was rendered on the order of 1/1,00,000 real-time.

...

and Toy Story before it

rendered by a few 75-100Mhz processors, real world effects of small caches/busses/ram/interconnects/etc means probably less than 50% realworld perf, with pathetic busses, memory, and ridiculous latency, and also given that ps2(2000h/w) had quite large processsors with 10M trans, and that was more than half a decade later(remember they took a long time doing it so they didn't start in 1995 but far far earlier 4yrs earlier)... I'd say their processors had around 1-2M trans... but I dunno...

ps3 should likely feature(between gpu & cpu) around 10-20times the amount of transistors of the ENTIRE RENDERFARM, at around 40-50times its speed(4Ghz estimates...).... with 1000s of times more bandwidth, far lower latency, and if my guess is correct could very well pack a few gigabytes of ram(mostly cheapo)... not only that but it LIKELY has h/w acceleration for some of those gphx calcs, making it 1000s of times faster at those calcs than the software rndring that was likely used here...

"Toy Story was the first fully computer-generated full-length feature film. The images required 800,000 hours generation time on a RenderFarm consisting of 87 2-CPU SparcStation 20's, 30 4-CPU Sparc-Station 20's and a SparcServer 1000. "

Wow!!! Look at all those cr@ppy pcs, with cr@ppy interconnects, and pathetic amounts of high latency, pathetically slow and small ram and hdds... This is gonna be owned badly...

"264 megabyte-per-second MBus Sun uses in its SPARCstation 10 and 20 serves not only as a memory path but as the computer's general system bus. " Remember that is THEORETICALLY most likely.

The cell in ps3 could very well have over 1000+ times the bandwidth with its embdded stuff...

PS
Of course it's not just ps3, but all consoles will likely feature similar perf(worst thing that can happen is a delay for whomever underestimated the perf of their competitors.).

HL2 gphx for next gen. is certainly not enough!!!

 

[KnightBreed]

zidane1strife, did you purposely make your post read like that or was it purely sarcasm?

For any interested party, here is a link to some common SPARCstation 20 configurations.

http://www.sunstuff.org/hardware/systems/sun4/sun4m/SPARCstation20/

ps3 should likely feature(between gpu & cpu) around 10-20times the amount of transistors of the ENTIRE RENDERFARM, at around 40-50times its speed(4Ghz estimates...).... with 1000s of times more bandwidth, far lower latency, and if my guess is correct could very well pack a few gigabytes of ram(mostly cheapo)... not only that but it LIKELY has h/w acceleration for some of those gphx calcs, making it 1000s of times faster at those calcs than the software rndring that was likely used here...

The hyperSPARC CPU used in those workstations had just shy of 1.2 million transistors, so 294 CPUs at 1.2 million a pop makes roughly 350 million transistors with 10 - 15 GB of RAM and ~500 GB of harddrive storage (guessing around 4 GB per workstation).

 

[zidane1strife]

The hyperSPARC CPU used in those workstations had just shy of 1.2 million transistors, so 294 CPUs at 1.2 million a pop makes roughly 350 million transistors with 10 - 15 GB of RAM and ~500 GB of harddrive storage (guessing around 4 GB per workstation).

So well, I overestimated a little, heh, didn't notice the 2 and 4 cpus numbers, well 2-3B trans, is about 5-10times the amount of trans of that at 40-50 times the speed.

In any case as with all projects, the higher-end machines were likely added later in the final yrs or months of the project... and thus their contribution is diluted by the low perf. of the first yrs.

10 - 15 GB of RAM and ~500 GB of harddrive storage (guessing around 4 GB per workstation).

Wow, for 1992-1993(first yrs, during which time they likely had less machines, etc, and contributed a lot to the 7hr per frame average.), that is more than I expected. But still, the busses are still quite slow, the amount of ram per machine quite small, the connections between'em quite pathetic.... etc.

ed

 

[KnightBreed]

Wow, for 1992-1993(first yrs, during which time they likely had less machines, etc, and contributed a lot to the 7hr per frame average.), that is more than I expected. But still, the busses are still quite slow, the amount of ram per machine quite small, the connections between'em quite pathetic.... etc.

ed

Not only would you have to condense the entire renderfarm into 1 or 2 microprocessors, but you would also have to increase performance 100,000x just to render the exact scene in realtime.

We could probably render a reasonable facsimile on the PS3 or GC2 or whatever, but corners would have to be cut, I'm sure.

 

[zidane1strife]

Rendering an individual frame of film could take anywhere from 45 minutes up to 20 hours to complete.
The 110 computers in Pixar's Renderfarm operate on a 24- hour basis.

So the lowr detail scenes only require (oops, heheh.) 61400 times the power it seems...

A team of 22 technical directors (TD's) were responsible for creating the film's 400+ models (which includes 76 characters).
All the models in the film comprise 4.5 million lines of code or 270 megabytes.

That fits nicely in any next-gen consoles memory.

Woody is described by 52,865 lines of model program.
He has 712 "avars" (animation controls) -- 212 in his face including 58 in his mouth alone.
There are 15 shaders and 26 texture maps on this character.

Well if we reduce the animation controls, he wouldn't be out of the question, now would he

edited

 

[zidane1strife]

Not only would you have to condense the entire renderfarm into 1 or 2 microprocessors, but you would also have to increase performance 100,000x just to render the exact scene in realtime.

Well, from what I've heard supposedly h/w acceleration can give orders of magnitude increase in speed for certain gphx calcs over sftwr rendering.

Furthermore as I've said, we've got likely over 5x the amounts of transistors of the renderfarm running at 40-50x the speed, and with decent interconnections/ram/bandwidth.

And as I said in the previous ati thread, and in this thread, the real world perf. of the rndr farm do to being distributed across an entire room instead of on a single chip, is most likely less than 50%, likely around 25%....

So cut your number to 25-50k, and cut it by 2-3orders of magnitude, due to h/w accel and talking about the lower complexity scenes, and you've got 25 to 500x ... now that is the number to beat, and with 40-50times the speed, and 5-10times the trans budget.... that doesn't seem to be entirely out of the question.

PS Althought, it won't be the same, I doubt 700points of animation are going to be given to chars, and 30+lights(most complex scenes) used, etc... next gen. making similar gphx more viable.

edii

 

[Simon F]

"Toy Story was the first fully computer-generated full-length feature film. The images required 800,000 hours generation time on a RenderFarm consisting of 87 2-CPU SparcStation 20's, 30 4-CPU Sparc-Station 20's and a SparcServer 1000. "

Wow!!! Look at all those cr@ppy pcs, with cr@ppy interconnects, and pathetic amounts of high latency, pathetically slow and small ram and hdds...

Funny, but I didn't see a single PC listed in the above.

This is gonna be owned badly...

Speak English.

 

[zidane1strife]

Funny, but I didn't see a single PC listed in the above.

thanks for the correction, with all the modern rndr farms using intel, and reading about intel's new gmr cpus, my mind is too hazy...

ed

 

[Gubbi]

zidane1strife, did you purposely make your post read like that or was it purely sarcasm?

For any interested party, here is a link to some common SPARCstation 20 configurations.

http://www.sunstuff.org/hardware/systems/sun4/sun4m/SPARCstation20/

<snip>
The hyperSPARC CPU used in those workstations had just shy of 1.2 million transistors, so 294 CPUs at 1.2 million a pop makes roughly 350 million transistors with 10 - 15 GB of RAM and ~500 GB of harddrive storage (guessing around 4 GB per workstation).

Wasn't the renderfarm sponsored by SUN? If it was I have a hard time believing they had HyperSparcs in them. HyperSparc was Ross' competing CPU, and usually sold as a two cpu module with level 2 cache on it. They are more likely to have had SuperSparc CPUs in them (also as an upgrade). SuperSparcs were about half the speed of HyperSparc (40-50 MHz 3-way superscalar vs. 80-100 MHz 2-way superscalar).

Anyway. There's still no way next gen. consoles will be 400.000 times faster than that renderfarm.

Cheers
Gubbi

 

[zidane1strife]

There's still no way next gen. consoles will be 400.000 times faster than that renderfarm.

For the lwr complexity scenes only 61,400 perf increase is required...(and remember it's not the THEORICAL number that has to get 61K increase, but the real-world which is likely less than 50%, probably 25%, the theoretic perf.)

ed

 

[Gubbi]

There's still no way next gen. consoles will be 400.000 times faster than that renderfarm.

For the lwr complexity scenes only 61,400 perf increase is required...

which isn't going to happen either.

Cheers
Gubbi

 

[zidane1strife]

Well, as I said if realworld perf is 25% it's only a 15K increase that is required and let's not forget

Pixar renders lighting test frames at a resolution of 720 by 432 pixels, requiring approximately 40 workstation-minutes per frame. Getting the lighting just right is difficult in any film, and a shot is often rendered several times before the lighting specialists are satisfied with the effects....

When a shot passes both animation and lighting tests, final rendering is done....

The same technology that makes Pixar's films so unique also presents them with the greatest technical challenge - the need for rendering power. Animation test, lighting test, and final rendering requires each frame to be rendered a minimum of three times.

All that most likely goes into the -mystical computer hrs- that are always cited... thus buffing'em up quite some... and we certainly won't have to render scenes several times with different adjustments before displaying'em. (some animation tests take as little as a few secs... but the rest do take quite some time.)

ediv

networking with 100 Mb/sec ethernet

Well one MB is 8Mb right?

so that's about 12.5MB/sec theoretical... so we're dealing with a b/w difference of likely more than 4 orders of magnitude compared to the inner cell bandwidth... and let's not forget the latency difference here.

Pixar requires seamless integration with its network of Apple®, DEC®, IBM®, SGI, and Sun machines. Rendering systems must connect to Pixar's FDDI backbone, support 10 Mb-per-second Ethernet, and allow expansion to 100 Mb-per-second Ethernet. With film production in process and on a tight schedule, Pixar cannot risk any disruption on its existing network.

So in most machines it's likely as low as 1.25MB/sec theoretic perf... that's quite pathetic...

The RenderFarm was expanded during the production process with an additional 120 processors -30 quad - processor SPARCstation 20 Model HS14MP systems - to enhance performance in speed lighting tests.

I knew IT!!! The highest perf ones were added later on.

RenderFarm uses a network computing architecture in which a powerful SPARCserver 1000 acting as a "texture server" supplies the necessary data to the many rendering client workstations needed to complete the rendering process.

interesting, wonder how those move through 1.25MB/sec(theoretical perf) connections

Each frame used up 300 megabytes of data

Certainly fits in ps3 memory... but alot of that data likely had trouble going through all those SLOW connections

 

[zidane1strife]

Cray-1. Cray's first supercomputer. Introduced
in 1976, this system had a peak performance of 133 megaflops.

...
Now pixar's toy story render-farm.

its total of 300 processors represents the equivalent of approximately 300 Cray 1 supercomputers.

39.9Gflops perf for the render farm at full Theoretical capacity.

Real-world, with very optimistic estimate for such slow/small connections/ram/hdd/caches around 10Gflops...

Prior to renderfarm expansion... probably 4-5Gflops.

Divide by optimistically low 4-5 due to the fact that scenes are rendered many a times... 1Gflops... and I doubt those are of any decent precision(but who knows.)

The ps3 could have a combined Tflops rating of 1-3Tflops, according to many... and it likely has h/w acceleration and textures/data are not going through 1.25MB/sec connections...

That is over 1000-3000 the perf of the renderfarm, and on top of that we have none of its bottlenecks(orders of magnitude more bandwidth.), and we likely have hw acceleration.

 

[Gubbi]

Assuming 4 floats per pixel a 300MB frame has a resolution of some ~18Mpixels or ~ 5,700 x 3250.

Considering the preview frames took 40 minutes to calculate and final renders were approx 60 times that resolution (assuming 4 floats per pixel) you'd be likely have a good deal more than 40 minutes to send those 300MB down the wire.

But even with 40 minutes, you'd only need 125KB/s (or 1Mbit/s) bandwidth to transmit 300MB.

Cheers
Gubbi

 

[zidane1strife]

Well I'm not sure but this is what I've got:

RESOLUTION:The resolution of a digital image refers to the number of pixels stored. For "Toy Story," the resolution is typically 1536 x 922 pixels.

 

[notAFanB]

RESOLUTION:The resolution of a digital image refers to the number of pixels stored. For "Toy Story," the resolution is typically 1536 x 922 pixels.

before or after supersampling?