Base settings, just resized output to 8k to create more tiles and better see how everything works. Unfortunately I’ve only managed to make 13 of them work despite checking all 20 in OpenCL settings. Is there any limit of GPU’s rendering one frame?
We don’t have a limit on the number of OpenCL devices, but this kind of setup with 20 GPUs is not something that we’ve tested. My first guess would be not enough tiles, but I guess you already tweak the tile size.
Yes, I’ve experimented with tiles size and quantity to both keep gpu’s busy and provide enough tiles for all of them. Test was not done natively but via software we develop which might be a problem as well, just wanted to make sure before conducting any other tests. I’ll experiment more with software and drivers, if that won’t help I’ll try to conduct test natively. Will provide more informations for further usage after that.
@JuanGea We develop app for gpu cloud computing and for alpha version we choosed fully working blender as our milestone. OpenCL is universal for both Nvidia and AMD, also AMD gpu’s are more preferable for this solution(due to Nvidia restrictions on using consumer cards as computing units on rendering farms). For test purposes I picked Nvidia cards because for now we don’t have that many AMD’s, however, at this point app is OpenCL based, so I was unable to use CUDA. Nodes are usually packed with 10 GPU’s per unit so without interference into hardware I cannot test CUDA nor OpenCL natively with more than 10 cards. I guess we’ll have to reconstruct node for that.
For information purposes so far:
for blender 2.8 commit 191b8951f7af everything works fine both on CUDA and OpenCL up to 10 gpu.