Patrick Elliott <sel### [at] rraznet> wrote:

> Ah.. Yeah, that could be it. Not that one couldn't sacrifice "some"
> precision to gain speed that way, but it wouldn't be the same result.

Essentially the GPU would be an alternative CPU. It would execute the same
raytracing algorithms as a CPU would.

> If you want to be picky, then yes. The point though is that between the
> two approximations, true raytracing is closest to reality, even if you
> have to round the edges of some things to make them, "not perfect
> boxes".

They both capture some aspects of the way light behaves. Neither of them are
noticably faithful to "real light". Many raytracers do not support anything
but the famous triangle soup. The box primitive is just another idealised
mathematical abstraction that needs work to resemble reality.

> And the overhead... Like I have told several people, what you
> can do in about 50 lines of POV-Ray code would take a 1.5MB mesh file,
> not including textures, on most GPU based systems.

The original suggestion was that it is possible to use a GPU as a fast
floating point processor, and a claim was made that POVRay would benefit. So
the GPU would run POVRay. There are lots of caveats (few of which derive
from the scanline heritage of GPUs). I personally can't see how it would be
worth the time, and think that the benefits are minimal compared to running
on the CPU, but it wasn't being suggested that some sort of super-GPU would
take POVRay on with scanline techniques, as far as I could tell.

These days I find myself using mainly meshes in POV (I don't do much/any
abstract work). Compactness is to be prized, but only if it actually
resembles what you want. I model planets with sphere primitives, starships
with meshes.

> From a purely practical standpoint, GPUs are not practical for photorealism,
> even if they are currently faster at producing results.

I don't know about that - scanline methods (grab-bag that they are) have
been the only commercially viable (extremely practical, no?) route to
photorealism for some time, and GPUs are probably the fastest
implementation of that technique available "by weight". They are only
getting more capable as time goes on.

I think it's clear that scanline will have to lose out to raytracing and its
derivatives in the end, but if it's good enough for ILM I think it's
probably not fair to call it impractical. It's sometimes said that pure
raytracing produces images that are too clean to be photorealistic. That's
got as much basis in fact, but doesn't reflect (yes) the potential of the
technique any more than, say, Quake does of scanlining.

Tom

Post a reply to this message