Real-Time Ray Tracing : Holy Grail or Fools’ Errand? *Partial Reconstruction*

And I expect ARM to kill x86 within 2 to 3 years and Intel to go bankrupt in the same timeframe. Next?

 

The quote was pretty general. It's not out of the question that we might see a smattering of ray-tracing-based effects in a few years, especially since Intel's gone and bought itself a developer.

On the ARM thing though, what is to stop either AMD or Nvidia from running a skunk works project to slap a few x86 or ARM cores (respectively) in a Larrabeesque solution (besides money, time, engineering resources, licensing )?

 

well it's not entirely unexpected since it was intel saying it. But it still seemed like a bold statement to me.

 

Not a bad idea, AMD doing an x86 version of Sun's Niagara could prove interesting to the company's bottom line, hopefully in a good way.

 

The MMX technology that would change the world since Pentium (TM) firstly appeared on the market. ?

 

Nvidia demos real-time GPU ray tracing at 1,920 x 1,080
nVIDIA's got a demo at Siggraph running on 4 Quadro GPUs (1GB each) in what they've dubbed the Quadro Plex 2100 D4 Visual Computing System (starting at only $11,000 to anyone interested lol)

nVIDIA: "the ray tracer shows linear scaling rendering of a highly complex, two-million polygon, anti-aliased automotive styling application."

 

Someone needs to use the same hardware and show what rasterising or a hybrid approach compares visually because those screenshots look very poor. Neat technology, definately has some uses, but as a mainstream consumer the end result is less than visually stimulating.

 

Hey, it's a start. At some point hardware will be tailored more towards it. The future will still probably be in hybrid solutions for a very long time (I find what Gran Turismo is doing in terms of creating hybrid lighting approaches - even if they're restricted so far to garage views and such - is interesting here).

 

I think it's good that nVidia does a demo like this.
Intel not only wanted people to believe that raytracing is the future, but also that Larrabee is what you need to enter that future (and some hinting at x86 to trick people even more into thinking that Larrabee is something conceptually different from other GPGPUs).
nVidia just shows the obvious: Larrabee isn't going to be the only raytracer in town.

 

Here are some words about this demo: NVIDIA Case Studies: Compute-Enabled Graphics , Ray Tracing & Ray Marching (page 15)

My question: How many rays per pixel?

 

I suppose you mean "on average"? Because there's way to many different types of pixel to answer this concisely with on number.

 

I'm on my PS3 now so it's a bit trickier to copy and paste links, but there's a link to a presentation demoing realtime raytracing of a recreation of Rome on a stack of Q21 Blades (I think totalling 6 TeraFlops). Looked pretty good.

EDIT: here

 

I have no real knowledge on this, but on another forum (and the custompc link) I got the impression that the nVidia demo is infact 'hybrid' - only reflections, refractions and shadows are raytraced, the rest normal rasterized stuff?

 

They had a demo like that too, but the main demo they showcased was raytracing everywhere. My understanding is that this was done, ironically, because the graphics/CUDA interop speed is not yet where it needs to be

 

http://www.caustic.com/
Ray-tracing accellerator for games and other applications.
What do you think?

 

Just how bad are modern GPUs at incoherent raytracing?

 

Cool thanks for the link... may I suggest we start a new thread for this though rather than reviving this dead one, which was specifically related to the B3D article?

 

I think for the last 10 years or so a small part of the academic raytracing world has been off in lala land ... especially when they try to build accelerators. What use is it to only accelerate stuff which GPUs can already handle decently, while not having an architecture flexible enough to accelerate the really hard stuff?

So it can accelerate primary rays with a couple of secondary rays, big whoop. How are you supposed to accelerate GI, caustics and soft shadows efficiently with just that? Backwards raytracing only avoids a very small part of the hacks necessary for good looking images.

You need a lot more logic, a lot more data structures and a lot more hacks to implement the hard stuff efficiently ... GPUs are at the point where they are flexible enough to allow you to do that. Even if they aren't good at irregular code, they do use ASIC design and processes technology far beyond what a start up with a low volume product could manage ... and can afford to sell at low margins.

In the end I don't think they will be either flexible enough or fast enough ... with Larrabee on the horizon they need to pull a miracle out of the hat to convince me otherwise ... 10 TFLOPs of programmable floating point or so for instance.

 

Multipass rendering for supersampling??? Seriously???

 

http://www.caustic.com/dev_intro.php



The interesting bit is that t his accellerator enhances the capabilities of AMD, Intel, and Nvidia's existing hardware. I think they have a winner with this one! I just hope Intel doesn't buy them and once thwart competition.