Concernant l'utilisation possible du GPU, je cite Stephan Reiter[1] :
> By the way: as far as I gathered, most of the raytracing stuff is CPU
> based. Are there any plans to load tasks to the GPU?
I actually tried that with a GeForce 7 in 2006 and gave a poster
presentation at a raytracing conference. But the idea didn't turn out so
well: With current GPU architectures there's simply too much performance
wasted when communicating CPU<->GPU, especially read-backs are problematic
(even with PCI-Express). A CPU-only implementation running on two cores
outperformed the graphics card by a factor of 4-5. Using the GPU as an
additional processing unit to the CPUs also yields a reduction in
performance due to the associated overhead.
autrement dit (traduction très approximative) que la communication entre le CPU et le GPU consomme trop de ressource par rapport à l'utilisation d'un cœur supplémentaire
[^] # Re: compléments
Posté par Gart Algar . En réponse au journal Raytracing temps réel. Évalué à 2.
> By the way: as far as I gathered, most of the raytracing stuff is CPU
> based. Are there any plans to load tasks to the GPU?
I actually tried that with a GeForce 7 in 2006 and gave a poster
presentation at a raytracing conference. But the idea didn't turn out so
well: With current GPU architectures there's simply too much performance
wasted when communicating CPU<->GPU, especially read-backs are problematic
(even with PCI-Express). A CPU-only implementation running on two cores
outperformed the graphics card by a factor of 4-5. Using the GPU as an
additional processing unit to the CPUs also yields a reduction in
performance due to the associated overhead.
autrement dit (traduction très approximative) que la communication entre le CPU et le GPU consomme trop de ressource par rapport à l'utilisation d'un cœur supplémentaire
[1]http://icculus.org/cgi-bin/ezmlm/ezmlm-cgi?50:mss:2290:jhhho(...)