For a research project, I have written a Tessellation Shader to interpolate solution points of a 3-Dimensional mathematical equation. Therefore I don't use any techniques for displacement, I merely use the tessellation hardware to interpolate a lot of data points in parallel. I want to write a paragraph about the potential speed benefits compared with trivial interpolation on a CPU, but I can't find any resources about how fast tessellation is in terms of "generating new points per second" or "subdividing triangles per second". Even more interesting would be "new points per GPU clock cycle". But all resources I find are from the video game perspective talking about FPS-performance, which is useless in my case, because first I tessellate my model only once and second I don't use any displacement techniques which are normally used to increase visual quality but will lower the performance again.

Obviously the performance depends on the actual GPU, but a single data sheet as an anchor point would already help a lot.

  • 2
    $\begingroup$ I don't know about your research, but if you're going to bother to talk about speed benefits, it would lend you some weight to talk about actual speed benefits, as measured by real benchmarks, rather than hypothetical "speed benefits" as measured by hardware stats. $\endgroup$ – Nicol Bolas Aug 29 '19 at 18:28
  • $\begingroup$ I guess you are right about that. The problem is, that I have one specific differential equation and results may vary for different equations. But I also don't know for sure, because the equation come from a mathematician while I am the programmer in this project. A selfmade benchmark would be my last option, but an indicator from hardware specs would be a good overview. $\endgroup$ – mkay Aug 30 '19 at 16:10

If you’re only tessellating the model once and never changing the density of the mesh then tessellation is probably a waste of time here—any measurements you make would include all the extra work the GPU has to do to generate new points / triangles, which is orthogonal to what it sounds like you’re trying to measure. Tessellating the mesh ahead of time (in your tool of choice) and doing the interpolation / evaluation in a vertex shader would probably give you a better benchmark against a CPU implementation working on the same mesh.

| improve this answer | |

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.