How does NVCaffe work on MultiGPUs?

[linnanwang]

Hello All,

Can anyone kindly explain me how does this caffe branch work on multiGPUs (at code level)? Thank you.

[lukeyeager]

I'll point you to the original Multi-GPU pull request, which has a large discussion about implementation:
BVLC#2114

Note that some changes were made before it was finally merged into BVLC (and later, then this fork):
BVLC#2903