Nvidia and Microsoft optimize Cognitive Toolkit for GPUs

Microsoft has partnered with Nvidia to optimize its Cognitive Toolkit for GPUs. The deep learning toolkit has applications in Cortana and Skype. The software should be approximately seven times faster than the previous version after an update.

For example, the toolkit can run on the Nvidia ‘supercomputer’ DGX-1 announced in April, which has eight Tesla P100 accelerators. Other possibilities are applications on Microsoft’s Azure servers, Nvidia writes. The Azure N VMs with Nvidia GPUs should be widely available soon and are now in a preview phase, the company said.

Ian Buck, Nvidia vice president for accelerated computing, explains to PCWorld that its own GPUs have already worked with the Cognitive Toolkit in the past, but that the collaboration has led to a large speed gain. To achieve this, the companies have jointly developed deep learning algorithms that are optimized for GPUs. Buck did not provide information about the availability of the new version of the toolkit.

Microsoft released the Cognitive Toolkit on GitHub in October and developers were able to combine it with Python and C++. The software was first known as CNTK and can be used for deep learning purposes. For example, the toolkit forms the basis for translations in Skype and finds out the context of searches in Bing. Microsoft applied the toolkit in its speech recognition technology, which recently recognized words at almost the same level as a human.