torok said:
It's similar to what you do in a regular desktop with CUDA. However, GPUs are heavily parallelized, while CPUs are more single-threade oriented. So what you can offload varies. If you are talking about computations full of conditions and different paths (AI), it's a pure CPU task. But some tasks are easily paralelized, like some types of physics. So you run them on the GPU. Also mind that GPGPU on PC is very different, because you don't have an unified memory. So if you are dealing with data on your CPU (system RAM) and wants to pass some task to the GPU, you have to copy it to the GPU's VRAM. That's very slow. An unified memory allows you to offload tasks that wouldn't be interesting on a shared memory architecture. So you are correct. A routine like physics that runs on the GPU would be faster on the Neo, so it would depend if the CPU tasks of the game can be parallel or not. |
Interesting stuff for sure. And as far as the unified memory pool goes, GDDR5 was chosen because of its high bandwidth correct? Going by the leaks then, It would be cost prohibitive to do a major CPU upgrade when they're supposedly increasing the throughout of the Ram, while the small incremental CPU upgrade may be "crutched" by the exponential increase in GPU via GPGPU Compute. That really puts 3rd parties in a sort of predicament regarding optimization. Do I spend the extra time coding for such a unified APU to hit certain targets, do I scale back certain assets, do I just make the title exclusive and focus solely on that architecture to reap the optimization benefits, or do I just say screw it and launch a bug ridden game.
In any case, Neo is sounding better and better. It's as if these consoles should be the real upgrade from 7th gen, and the X1/PS4 are the ".5's"







