Performance Formula

[ghiom3d]

Hello,
It's a question for expert about performance setup in RT GPU.

I would like to know if it's possible to find a " performance formula " to put correct numbers in RAY BUNDLE SIZE and RAYS per PIXEL .

I've try to understand how it's work but it's to hard to me .
i have 6 gpu , if i grow up RAY BUNDLE SIZE from 256 to 512, i have a better performance. if i try more than 1024 the performance down in rendertime, even it's growing in the stats .
If i try less than 256, rendertime are a little slower and the refreshing is better.
...

so i thing that is in relation about how much cores have GPU and the size of the rendered picture. And maye how many GPU we have .

Can we find a math operation to setup this parameter ?
for example, in the test GPU scene, i've find that 512 for ray bundle size looks better than 256 and i save 5 seconds for 5000 path/pixels reached .
With 640 and more rendertimes are longer .
360 , longer rendertime.
???

Any ideas, helps, explanation ?

[vlado]

You can also try different rays/pixel values. But in general you will have to experiment a little; if there was one ideal setting, we would have just hard-coded it

Best regards,
Vlado

[ghiom3d]

thanks for your answer, but can you explain what each value mean ?
does the ray bundle size is something like bucket size ?
what it's mean ?

[vlado]

The bundle size is how many pixels are sent to the GPU for processing in one batch. Once these are computed, the results are read back by the CPU. The samples per pixel determine how many paths will be traced for each of these pixels.

Low bundle size and samples per pixel mean that the GPU does less work for each batch, so it can process it faster. However its usage may drop as it needs to wait for the next batch to come along, even if the wait time is quite low.

Larger bundle size means that more work is given to the GPU, so that it works efficiently for a longer period of time. Larger bundle size also means that more data is transferred to and from the GPU, which may slow things down for larger values.

More samples per pixel mean that again the GPU is given more work, however this does not have as great impact on the amount of data transferred from the GPU as increasing the bundle size, since only the final pixel result is read back by the CPU, and not the individual sub-pixel samples. If increasing the bundle size does not improve performance, then you can try increasing this value instead.

Best regards,
Vlado

[3LP]

Nice,

Thanks vlado

[ghiom3d]

It's interesting .. thanks you Vlado !
I understand now

[Alan Iglesias]

It might also be appropriate to mention here that changing the Bundle Size and Samples per Pixel can greatly affect your level of interactivity if you are using your display card to help with the GPU rendering like I currently am. I've been running them lower during scene creation, and am very happy to see the rendering times not really affected while getting much better interactivity in the interface, which I think can be very useful and a lot of fun. This is with a single machine with two cards. But as has been mentioned, when using multiple machines/GPUs the BS and S/P amounts can be increased for even faster rendering.

I guess the bottom line is that there will be plenty of experimentation going on as we find what is best with our particular system/application. Thanks for the excellent explanation, Vlado.

-Alan