|
|
In article <cja### [at] netplexaussieorg> ,
Christopher James Huff <cja### [at] earthlinknet> wrote:
> Still? Meaning they've used it before?
Yes, since August. I had pointed out to them in detail that using
optics.pov is problematic (and got a response), but of course only after
they published first results in the print issue. So it was too late for
them to change. I suppose this is one reason why they now report results
for POV-Ray 3.1 as well as POV-Ray 3.5 again (in August they only reported
3.5 results). As you know, the problem with optics.pov is that is spends a
lot of time in photon code, which is more searching than FPU usage.
> It is definitely not a good general benchmark scene, its purpose is to
Very true.
> demonstrate media photons. How could they miss the new benchmark scene,
> and what could make them choose optics.pov?
I don't know. I guess they were afraid POV-Ray was specially optimized for
benchmark.pov, which to some extend is true (also not for specific
processors per-se), but it is still a better choice. And optics.pov takes
reasonably long, so to someone who does not know the implementation details
it may seem like a good benchmark scene :-(
However, I have to admit that even when putting all the optimisation aside,
the raw benchmark.pov isn't a perfect benchmark as I found after doing some
extensive profiling. It turns out a very few object types dominate the
tracing process, which creates a bias and only evaluates a small amount of
code. It can be fixed*, but then the scene looks really terrible...
Thorsten
* I will post details elsewhere.
____________________________________________________
Thorsten Froehlich, Duisburg, Germany
e-mail: tho### [at] trfde
Visit POV-Ray on the web: http://mac.povray.org
Post a reply to this message
|
|