NVIDIA doubles down on its artificial neural network development efforts
NVIDIA has doubled down on its artificial neural network development efforts. Read on to find out more.
DIGITS console (Image source: NVIDIA)
With the announcement of DIGITS Deep Learning GPU Training System version 2 (DIGITS 2) and CUDA Deep Neural Network library version 3 (cuDNN 3), NVIDIA has deliver new software to enable data scientists and researchers to boost their efforts on their respective deep neural network projects. DIGITS 2 is the updated version of DIGITS, which was first announced in March at GTC 2015, the main improvement of DIGITS 2 is how the training system allows automatic multi-GPU scaling.
Deep learning neural network development and deployment workflow process (Image source: NVIDIA)
As the DIGITS system features a GUI, the latest version allows the data scientists to distribute the deep learning training workload across multiple GPUs. According to the company, with DIGITS 2, their engineers have “trained the well-known AlexNet neural network model more than two times faster on four NVIDIA Maxwell architecture-based GPUs, compared to a single GPU.”
For data researchers, NVIDIA has also announced the CUDA-based cuDNN 3 library. In comparison with the previous version, cuDNN 3 is touted to train deep neural networks up to 2x faster on a single GPU. cuDNN is a library of mathematical routines for deep neural networks. Developers integrate the library routines into the learning framework of higher-level machines. cuDNN 3 features support for 16-bit floating point data storage in the GPU memory.
This allows for larger and more complex neural networks due to the increment (two times more) in amount of data storage, as well as the optimization of memory bandwidth. According to a Baidu researcher, cuDNN 3 also manages to improve the accuracy of their artificial neural networks. The preview for DIGITS 2 is available now for registered NVIDIA developers; however, in order to drive this training system, you would most likely have to invest in a US$15,000 DIGITS DevBox, which is deskside deep learning machine. The cuDNN 3 library is expected to be integrated into major deep learning frameworks, which include Caffe, Minerva, Theano and Torch, in the near future.
Our articles may contain affiliate links. If you buy through these links, we may earn a small commission.