How much does the performance of cuda applications depend on the number of cuda cores?

Question

GPGPU

With all other characteristics being equal for video cards, a twofold difference in the number of cuda cores will give a twofold difference in the performance of cuda applications? I choose a video card for unbiased rendering on gpu, if it matters.

Answer 1

A twofold increase will be only with infinite memory bandwidth. If it is not possible to fully load the gpu, the number of cores will not give an effect. It is necessary that the costs of data processing exceed the costs of reading and writing them to memory so that the execution time of the latter can be neglected.

Answer 2

In an ideal situation - a linear relationship, the ideal situation depends on the speed of external processing of this data: how quickly they will be loaded and unloaded from memory, how quickly new data will be fed in, how long already processed data will be processed on the CPU.

Answer 3

I have not tried it with rendering, but it increases the math quite proportionally. If there are no bottlenecks, then it should also be proportional or so.

Answer 4

It all depends on what the performance of the application rests on. If this is exactly the math, then there will be a proportional (in theory) increase in performance, provided that this is the same architecture and the cores operate at the same frequency. But if the application rests on memory, then there will be no sense, since the cores will not have enough data and they will be idle. It all depends on the implementation of the renderer.

Answer 5

Tried iray on professional NV Quadro. I did not count the number of cores on the video card and did not time it, but the results indicated in this test are approximately true. For myself, I concluded that so far unbiased renderers are not the same.