Simply put, every one of the requests we see for a GPU accelerated 
POV-Ray is a case of a solution looking for a problem.  Unfortunately, 
it's the wrong solution.  In that regard, it's more of a hammer looking 
for nails.

GPU acceleration will be useful when the following conditions are met:
1) Support for sophisticated branching
2) Full double-precision accuracy
3) Large memory sets (other than textures)
4) Independent shaders running on distinct units.

The fact is that no GPU currently in production satisfies a single one 
of these critera.  Some are getting closer (for instance, both the 
nVidia 2x0 series and the AMD 48x0 series have much better support for 
double precision), but none of them have it.

Additionally, there are the following two criteria:

5) A standard interface for boards from multiple vendors
6) Must be fast enough to be useful

OpenCL *might* help - there's no way to know for sure until we see some 
implementations, and experiment with it.

The speed, on the other hand, is the real killer.  TMPGEnc was recently 
modified to utilize nVidia's CUDA, and with disastrous results.  A 
high-end nVidia card, the 260 I believe, was compared to a 2.4gHz Intel 
C2Q.  It turns out that the conversion ran *faster* with CUDA turned 
off.  So much for helping out.

Once these conditions are met, *then* the idea will merit another 
consideration.  Until then, it's just a waste of time.

-- 
...Chambers
www.pacificwebguy.com

Post a reply to this message

I think the wrong assumption here is that there's much of a future for
raytracing on GPUs except for small specialized details as reflections on cars
in racing games.

Post a reply to this message

Chambers <ben### [at] pacificwebguycom> wrote:
> Simply put, every one of the requests we see for a GPU accelerated
> POV-Ray is a case of a solution looking for a problem.  Unfortunately,
> it's the wrong solution.  In that regard, it's more of a hammer looking
> for nails.

Are you sure you don't mean a screwdriver looking for nails?

I think it's often the case that performance is a convenient substitute for
content.  It seems to me that the payoff of improving algorithms and features
would be much greater than that of going for sheer performance.

> Once these conditions are met, *then* the idea will merit another
> consideration.  Until then, it's just a waste of time.

It's an interesting dilemma.  Just think of the huge amount of resources being
poured into developing codes for computers like Roadrunner, and the processors,
standards, and languages that will replace it in a couple years.  Of course this
is the only way to move forward, but I think I'll wait a few years before diving
into GPGPU.

 - Ricky

Post a reply to this message