I'm no expert, but I think that a lot of the GPU speedup has come through massively parallel execution rather than processor/bus/memory speed. NVidia's Quadro TRX 8000, for example, has 4,608 CUDA cores, 576 Tensor cores, and 72 RT cores. I don't know what all those do, but they are certainly doing a lot of it in parallel!
Re:PlayCanvas launches glTF 2.0 Viewer Tool
by CyberDildonics on Saturday August 01, 2020 @08:29AM
(#24019963)
First, massively parallel execution takes more transistors and moore's law is about more transistors. Second, memory bandwidth has increased significantly.
Realistic rendering has not solidified and in any way. It is arguably in a much greater state of flux and advancement these days with lots of useful research coming from game and game engine companies. Not only that, it is unlikely that it would buy much speed to hardcode some model into a GPU and those transistors could be used instead for flexible computation. It would need to do lots of floating point math, which is already what shaders are great at. Shaders were invented to 'shade' fragments, that's why they are called shaders. They have been gradually adopted as kernels that are run on other things.