The demo (ZF) with cuDNN is a little slower than without cuDNN?

[JohnnyY8]

I have gotten the commit: 96dc9f1 according to #108 . But the cuDNN still slow the demo down.
Anyone can help me? Thank you!

[happyharrycn]

There is no guarantee that cuDNN will always be faster. The runtime depends on the network structure, your GPU model, the cuDNN version and how you choose to compute convolution within cuDNN. For example, cuDNN is considerably slower on VGG16 than Caffe's native implementation before V4.

[JohnnyY8]

Oh, I see! It is my mistake to think the cuDNN will always be faster. Thanks for your answer, it is really helpful for us.
But do you have any better ways to make sure the cuDNN is suitable for our model or other things? I always try with and without cuDNN respectively to compare. @happyharrycn

[happyharrycn]

Well, this is tricky question. The major advantage of cuDNN over Caffe's native implementation part is the convolution part. I was checking the code in Caffe and realized that Caffe is not using the full power of cuDNN. For convolution, cuDNN offers different algorithms with an auto-tune function that selects the best one. In Caffe, this auto-tune feature is "off" due to a management issue. You can also check the discussion here.

[JohnnyY8]

Alright, that looks really complex. But it is not necessary for studies.
Anyway, thanks for everything! @happyharrycn

[isalirezag]

@JohnnyY8 how can you disable using cudnn?