bbeepp wrote:I did a quick benchmark from demo scenes. PT kernel:
1.55 - 6.1 Ms/sec
2.0 - 3.4 Ms/sec
Done on single GTX 780.
Why there is almost 50% slowdown?
When I run the demo scene on 1 690 GPU, I get these speeds:
1.55: 3.74 Ms/s
2.00: 2.77 Ms/s
That's a slow-down of ~25%. It's not great, but not 45%. I really don't understand how you got to your numbers.
Just to make sure that we are rendering the same thing, that's how the trench looks like, when opened untouched in 1.55:
You may notice that the sun position has changed
To the why: We had to make a lot of changes in the ray-tracing and geometry code, to make displacement, motion blur, hair and rounded edges work. They introduce more complexity and therefor slow-down ray-tracing.
I can't tell you exactly why it's so much slower, because you can't profile CUDA kernels easily, but my running theory is: Since this particular scene mostly tests ray-tracing only (since the textures are trivial), it gets affected more by the slow-down in ray-tracing than scenes with more complex materials. We also had to modify the geometry compilation, so it's possible that we introduced a bug here.
All I can say is that we are currently investigating the performance issues. Unfortunately I can't make any promises that we can solve them very quickly.
In theory there is no difference between theory and practice. In practice there is. - Yogi Berra