Low Pascal render speed (Super speed starts after 5000 samp)

[coilbook]

We have master (old titan and 6 x 780TIs) + slave with 5 GPUs (2x980, 780TI, 1080 and new GTX Titan X (pascal) we get 365 Samples/s.

IF we disable slave and only use master (old titan and 6 x 780TIs) we get 275 Samples/s

So (2x980, 780TI, 1080 and new GTX Titan X (pascal) only give us extra 90 Samples/S Why so low???? Two of those cards are 1080 and new Titan X Pascal

[coilbook]

EDIT:
When rendering gets to 5000 samples out of 16000, samples/s increase to 578. is this some kind of super boost?

Why dont we get these kind of speed when image is at 5000 or less?

We do animation so we usually set our scenes to around 3000 samples. That means we cannot have full potential of those cards because super boost starts at 5000. And only 0.5 Gb of memory is used so the scene was loaded way before 5000 samples. Please help


p.s slave is one PC with netstor turbo box. PC has 980 and 780Ti and turbobox has 1080, new titan x and 980.. Maybe it does take a while for data to transfer from turbo box to slave and then to master

but if it is a load delay why it says Net GPU 5/5 all slaves are loaded and rendering.

[Sportler]

EDIT:
When rendering gets to 5000 samples out of 16000, samples/s increase to 578. is this some kind of super boost?

Why dont we get these kind of speed when image is at 5000 or less?

We do animation so we usually set our scenes to around 3000 samples. That means we cannot have full potential of those cards because super boost starts at 5000. And only 0.5 Gb of memory is used so the scene was loaded way before 5000 samples. Please help


p.s slave is one PC with netstor turbo box. PC has 980 and 780Ti and turbobox has 1080, new titan x and 980.. Maybe it does take a while for data to transfer from turbo box to slave and then to master

but if it is a load delay why it says Net GPU 5/5 all slaves are loaded and rendering.

My guess it has something to do with this part of the release information:

Experimental Pascal GPU support

We could narrow down the issue using the CUDA 8 RC toolkit and were able to work around it by doing a mixed CUDA 7.5/8 build for Pascal GPUs. For previous architectures we still use CUDA 7.5 only. This is all experimental and we haven't done any specific Pascal tweaks/optimizations yet and the performance isn't where we would expect it to be, but at least you can use your Pascal GPUs now.

The performance for Pascal cards is not yet 100% because Nvidia is yet to release the full CUDA 8 toolkit, and until then the support for Pascal GPU's are a little wonky.

[bepeg4d]

Hi coilbook,
have you tried to disable the "Minimize Net Traffic" option in the kernel settings?
https://docs.otoy.com/manuals/products/ ... ht-kernel/
ciao beppe

[coilbook]

Hi coilbook,
have you tried to disable the "Minimize Net Traffic" option in the kernel settings?
https://docs.otoy.com/manuals/products/ ... ht-kernel/
ciao beppe

Thank you

We tried this this morning it goes to 660 samples/s quick especially when minimize traffic is unchecked