- Fri Mar 10, 2017 11:55 am
This is almost philosophical. We are working on multi-gpu right now. Let's see how does really work in performance. The question is more when than what. GPUs are the future of any serious computation task. But program for GPUs well done is really heavy stuff. So we still need good CPUs right now. And we are working on some CPU engine developing too. Can we in the future do hybrid rendering? Things like CPU calculating heavy paths like caustics, and GPU running very fast the rest? Sounds very interesting, not really near. Other options are having compatible engines on CPU and GPU, so we can tile the image and calculate parts of it across different devices (of any kind, CPU or GPU). And even, for batches or animations, having each device to work on a different frame at the same time. But all this are just things we have though can be done, nothing serious on the table yet, we still have a lot to do on the GPU engine, and the whole maxwell environment. We can even talk about if CPUs and GPUs are kind of being more similar, with GPUs dealing better with bifurcations than before, and CPUs having more and more cores each time.
Anyway, if really using the machine for production, I would put a good CPU, and a decent GPU (anything of the series 1000, don't need to be the latest really expensive card). Anyway, dual systems, and systems over 80 threads seems to be giving worse performance for rendering than having a network of cheaper computers working on cooperative.