Auto-tune a convolutional network for NVIDIA GPU

#1

I want to auto-tune a network using 8 1080, but when i query rpc tracker, only 1 gpu was find.
image

#2

image
image.png669×638 127 KB

#3

by default the rpc_server uses gpu(0). You can start multiple of the rpc_server and set env variable CUDA_VISIBLE_DEVICES to different value

#4

It is ok now. thank you so much!

#5

how you set CUDA_VISIBLE_DEVICES to different value? I start multiple of the rpc_server but only one is used when tuning.

#6

You need to start an rpc server for each gpu with different env var