Follow the previous post to install, run on several machines, the operation speed is simple to remember.
1. Win7, i7-4790k, 32G (8*4)
Pure CPU calculation, first time pre-training per step: 42.x seconds
2. Y480 Notebook PC, 4G
GPU accelerated calculation, first time pre-training per step: 90.x seconds
[Resolution of new graphics card support issues]
After purchasing the shadow GTX960 4G, the runtime prompt error:
NVCC fatal:value ' sm_52 ' isn't defined for option ' gpu-architecture '
Online Search found the answer:
http://rexdouglass.com/training-neural-networks-on-the-gpu-with-commodity-hardware-installation-and-configuration/
Just add the following in the. Theanorc settings file:
[NVCC]
Flags =-arch=sm_30
The reason is that the CUDA6.5 version is too low, Katai, said to be 7.0 or 7.5 can, I am too troublesome, did not change.
Maybe the 7.x version can improve speed? Have time to try again later.
Then experiment on two desktops:
3. Win7, i5-3450, 8G (8*1)
GPU accelerated calculation, first time pre-training per step: 19.x seconds
4. Win7, i7-4790k, 32G (8*4)
GPU accelerated calculation, first time pre-training per step: 15.x seconds
Summarize:
1. i7-4790k is very powerful
2. Y480 laptop computer GPU is weak
3. The i7 desktop is faster than the i5 desktop, possibly because of the fast memory and the I7 dual channel
CUDA6.5 + GTX960 + winpython, running Theano DBN