Quote from: pokoy on June 05, 2019, 07:56:49 AM
As for CPU vs GPU, I guess porting RTP to GPU is quite an effort and it could in fact end up slower than CPU once it exceeds the GPU's memory. Also, supporting all the different GPU code ecosystems (OpenCL vs CUDA) and models is not trivial, and likely impossible for small dev teams.
That's a pretty bold statement. GPUs, like CPUs, have logic for communication. OpenCL vs CUDA is just a choice. One is proprietary, one is not. They both communicate with the GPU. The difference here is CUDA can use multiple GPUs, OpenCL cannot (it doesn't know how).
Additionally, like any GPU/Game/Renderer, it doesn't need to rely on only it's GPU memory.
Which todays, coupled with it's GPU, is much more efficient than a CPU by leaps and bounds with not only iteration speed but memory usage, because it has dedicated on board memory, and the systems on board memory. Most GPU units people use for rendering are at 6-8gb coupled with 16-32gb RAM in the machine. Workstation GPUs come with far more ram such as the Vega coming out with 32gb for Apple's workstation.
For the work TG is doing, especially moving to PT, it only makes sense to move to GPU rendering like the big boys. CPU rendering at home was really a means to fill a void that SGI made with it's GPU rendering that wasn't available to consumers at home.
------------
I'm not sure what's with the RTP on my CPU than. It is absolutely worthless and doesn't follow the programs System Requirements of a Quad Core. My system is 3.7ghz base, at 4.2ghz overclocked. Should be fine. For example the "proudly CPU" based Corona is very stable and quick on my system.
GPU rendering on my RX 480 8GB however is vastly quicker than Corona or TG, and in many cases for far more detail in objects and geometry.