How to accelerate 3D convolution by CUDNN V5

I use TITAN Black, the deep learning framework is caffe. I upgrade my CUDNN to V5, but it seem that my 3D convolution does not faster than before.

By the way, I use cudnnSetConvolutionNdDescriptor to implement the 3D convolution.

What should I do to accelerate 3D convolution by CUDNN V5?

Same question here. also wondering what framwork you are using to do 3d convolution.

Thanks!

Attic

I guess doing video analysis instead of image analysis would require 3d convolution.