There are programming libraries for almost every conceivable thing out there and deep neural networks have not been left out. The NVIDIA CUDA Deep Neural Network library or cuDNN is one such library that comes with a host of benefits.
cuDNN is part of the NVIDIA Deep Learning SDK provides implementations of standard functions for some of the functions areas such as pooling, normalization, activation layers, forward and backward convolution and more.
So What are the Benefits of NVIDIA cuDNN?
No developer or researcher wants to spend unnecessary time tuning or optimizing a system when they need not be doing so. They are more interested in focusing on the actual work and getting results for their work.
cuDNN helps in that it takes care of the bare bones stuff thereby freeing the users to work on more important parts of their research such as training neural networks and developing software applications.
cuDNN is capable of accelerating several popular deep learning frameworks such as Theano, TensorFlow, Caffe, Torch, and CNTK among others. You can get a comprehensive list of supported frameworks here.
The most up-to-date version available for download is cuDNN 5 Release Candidate. Over previous versions, this release delivers quite a few benefits as highlighted below.
- On a single NVIDIA Pascal GPU, you can get over 40% fast training of the deep neural network
- You get up to 6 times more speed in Torch on LSTM recurrent neural networks
- You can now have accelerated networks with additional routines such as 3×3 convolutions, such as VGG, GoogleNet, and ResNets
- Jetson TX1 now has support for cuDNN 5 and above releases
- On Pascal GPUs, you now get reduced memory usage with FP16 routines and subsequently increased performance
The NVIDIA cuDNN is supported on Windows, Linux and Mac systems with Pascal, Kepler, Maxwell, Tegra K1 or Tegra X1 GPUs.
To get the free NVIDIA cuDNN one would have to register and become a member of the Accelerated Computing Developer Program.
Ref:
Further Reading: https://developer.nvidia.com/cudnn
Found this article interesting? Follow Brightwhiz on Facebook, Twitter, and YouTube to read and watch more content we post.
