Which graphics quality to expect from a supercomputer like Earth Simulator?

[Benedikt]

Hi there,

I always wanted to know which realtime graphics quality you could expect from a supercomputer like the new Earth Simulator in Japan? Let's say, you could buy one for at home...
If you compare the mflops/s of a modern pc and the Gflops/s or Tflops/s (?) of a supercomputer, could you estimate the graphics power of a supercomputer like that?
I am just interested in which kind of games or simulations you could run on a caliber like that. UT 2003 in 2048x1536 with 16 AA and so on .... Or could it be real photorealistic graphics like the newest blockbusters like Star Wars 2? In realtime of course, with YOU in the middle....

---> and this leads to another question:

What about real time raytracing? Let's say, a "on-the-fly-generated" render movie with a constant rate of 25fps, and full interactivity. And couldn't this be the future ? That you change from the polygon-based graphics models of today to 3D scenes generated by raytracing? Real 25fps of computer-generated, interactive raytraced video!
I think the power should be there shouldn't it be?

Cu and thanks in advance...

Greetings!

BW

 

[CTho9305]

It would not necessarily measure the way you'd think, because supercomputesr are very parallel. If you had a supercomputer just increment a number, it probably woulnd't be more than 2-3 times faster than a normal comptuer (because it is made up of individual processors that are often faster than consumer chips). For rendering pixels (a VERY parallel task), however, if you had enough bandwidth between them, it would be VERY fast.

 

[BurnItDwn]

just use each node to render a specific line of detail ... or split it up like the SLI Voodoo 2 cards ... but on a much larger scale ....

I would imagine it to be fully capable of real time raytracing ... however to get it working would be the hard part

 

[everman]

Well...they aren't exactly designed for that kind of thing. I suppose it is possible though.

 

[Howard]

12-bit color is a necessity if you want lifelike models...

 

[Shalmanese]

Originally posted by: Howard
12-bit color is a necessity if you want lifelike models...

the radeon 9700 can alreadt do 128 bit colour and you want 12 bit?

 

[Sunner]

Originally posted by: Shalmanese

Originally posted by: Howard
12-bit color is a necessity if you want lifelike models...

the radeon 9700 can alreadt do 128 bit colour and you want 12 bit?

Maybe he meant 12 per channel?

 

[Howard]

Yeah, that's what I meant.

Imagine watching The Spirits Within in 16 colors.

 

[Sahakiel]

Originally posted by: Howard
Yeah, that's what I meant.

Imagine watching The Spirits Within in 16 colors.

Landscapes and textures may be fine and dandy, but I'm really itching to see full-motion realtime rendering with realtime physics, musculature, and the like. Imagine watching realtime video rendered so well you can't tell it's all virtual.

 

[dejitaru]

Landscapes and textures may be fine and dandy, but I'm really itching to see full-motion realtime rendering with realtime physics, musculature, and the like. Imagine watching realtime video rendered so well you can't tell it's all virtual.

There were some rendered videos on the Discovery channel a while ago which looked not realistic, yet real (for a few seconds).
It isn't often I look at a landscape, tree, or anything without wondering how long it would take to render.
Realistic modeling, animation, and rendering are needed for a "real" animation, all of which put a nice strain on yer math chips.

I am just interested in which kind of games or simulations you could run on a caliber like that. UT 2003 in 2048x1536 with 16 AA and so on

Your Radeon 9700 would be a nice bottleneck, though the computer would be able to run it quite well without any hardware "acceleration."

Imagine infinitely fast processors.

 

[Howard]

Software mode, then?

 

[Ben50]

I guarantee that if you put a computer like that in the hands of ILM or the like, they could make the most realistic graphics you've ever seen.

 

[wfbberzerker]

i remember reading about some graphics-type supercomputer (no it wasnt like the renderfarm, it did it in realtime).
i think it was capable of doing something like 90 million triangles/sec

 

[FuriousBroccoli]

Originally posted by: wfbberzerker
i remember reading about some graphics-type supercomputer (no it wasnt like the renderfarm, it did it in realtime).
i think it was capable of doing something like 90 million triangles/sec

Umm the Radeon 9700 can do 350 million triangles/sec.

As to the original question. Assuming writing the software to render it wasn't a problem, your bottleneck isn't going to be the computer, it'll be the artists! If you want to see what's possible with future technology and todays artists, just look at any offline render..people are prepared to wait yonks to render an image already, a supercomputer would just make that job faster, not necessarily better.

BTW, We don't need raytracing, pretty much everything can be done without it. It won't replace polygons for at least a few decades, and perhaps not ever with the other growing techniques that can fill the void.

 

[dejitaru]

polygons == bad

They need to be replaced.

 

[FuriousBroccoli]

Originally posted by: dejitaru
polygons == bad

They need to be replaced.

Polygons aren't bad, they are easy to work with and can scale to any detail level. Mind sharing what you think we should be replacing them with?

 

[Shalmanese]

fractal polygons!

Or at least 2nd degree polynomials (like what ATI was doing with truform except in the design stage, not the redering.)

 

[silverpig]

Check out SGI as they have some supercomputers for graphics.

 

[dejitaru]

Polygons aren't bad, they are easy to work with and can scale to any detail level. Mind sharing what you think we should be replacing them with?

Geh, well, polygons are ok I suppose.
We need polyhedrons. The Xbox and such have accomplished pretty much all there is to do with polygons (on a practical level), yet a completely new rendering system will not likely be accepted any time soon. A hardware accelerated raytracer would have been nice.
Polygons are a hack. As are textures- and those damn "pixels"

Perhaps I'm looking too far ahead.

 

[silverpig]

What SGI does is they composite a number of high resolution still images taken at different angles together to come up with a 3d object. No polys for many of their models. Check it out on that site I posted... It's in the infinitereality4 section.

 

[Shalmanese]

Arent polyhedra just a lot of polygons? How do they differ from polygons? unless you are implying stuff like 3d textures.

I still dont see a way to get rid of pixels unless you want to fundamentally change monitor technology as well as graphics technology.

 

[dejitaru]

Arent polyhedra just a lot of polygons? How do they differ from polygons? unless you are implying stuff like 3d textures.

With polyhedrons you can make hypercubes and 4D models. Yet these, alas, cannot be usefully represented on just a 2D computer monitor.

I still dont see a way to get rid of pixels unless you want to fundamentally change monitor technology as well as graphics technology.

Yes. Yes, I do. Right now.

 

[FuriousBroccoli]

Originally posted by: Shalmanese
Arent polyhedra just a lot of polygons? How do they differ from polygons? unless you are implying stuff like 3d textures.

I still dont see a way to get rid of pixels unless you want to fundamentally change monitor technology as well as graphics technology.

A lot of people are naively idealistic when thinknig about rendering. Of course polygons are a hack, but there's no such thing as a non-hack. You can never EVER model anything as complex than the machine modeling it (to do so would violate the rules of the universe, and common sense -- if a CPU could simulate a faster CPU, then you have created an infinitely powerful system by the nature of recursion), so the only question is what are the advantages of different hacks.

At the end of the line you *MUST* have a *discrete* approximation of whatever you're rendering, and polygons and pixels are the simplest discrete approximations. As long as polygons are sub-pixel in size (and they will be very soon in realtime), then you CAN NOT improve on the representation.

Same for raytracing. It's an approximation again, and a terribly slow one, and until we have infinite processing power (hint: never), we need to make tradeoffs. We won't get raytracing hardware, because games are increasing in complexity as fast as processing power is increasing, raytracing by its nature is going to get exponentially complicated with the scene, so really is very unlikely to become practical.

There's of course the question of primitives for modeling, and that's a totally different area. We already have mathematical methods at artists disposals, such as NPatches. They have many disadvantages though, and they can be very hard to work with.

So the question I ask to anyone who wants to scrap polygons, is what is your goal? To make artists lives easier? To have a mathematical way that can scale in detail to match hardware? To make things faster for the hardware? To reduce memory? Because polygons can do anything any other primitive can, it's just all about relative merits.

 

[dejitaru]

So the question I ask to anyone who wants to scrap polygons, is what is your goal? To make artists lives easier? To have a mathematical way that can scale in detail to match hardware? To make things faster for the hardware? To reduce memory?

Yes, yes, yes, and yes. You raise good points, but I still hate polygons. :| No technology is utilized forever.