Hi,

I tried to use caffe to train an imagenet model. But the training speed is 2~3 times slower than the benchmark. I am using Cent OS 6.5, Tesla K40m. I have faced the same problem with a Black TITAN on Ubuntu 12.04, which can be addressed by changing the Nvidia driver.

The speed for training MNIST and CIFAR10 demo is normal. Is there anyone who are able to get the benchmark speed for imagenet with CentOS 6.5 and K40?

Any help will be appreciated.
Dennis