Microsoft and Nvidia have teamed up to push AI on premises and in the cloud.
A deal between Microsoft and Nvidia has resulted in Microsoft's Cognitive Toolkit being optimised for Nvidia's DGX-1 supercomputer in a box, and the installation of Nvidia GPUs in Azure data centres.
Microsoft's Cognitive Toolkit is the technology behind the Cortana assistant, the Skype language translator, the Bing search engine, and Hololens augmented reality.
Cognitive Toolkit has been optimised to run on systems using Nvidia's Tesla GPUs. Where a conventional dual-socket server running AlexNet on the Cognitive Toolkit can train 78 images per second, a single Tesla P100 can do 2,400 images per second, according to Nvidia vice-president of solutions architecture and engineering Marc Hamilton.
{loadposition stephen08}Two P100s pushes that to 3,500, and four to 7,600. And a DGX-1 with eight P100s can achieve 13,000 images per second.
The Cognitive Toolkit with Nvidia support is available for use on premises or in the Azure cloud. In the latter case, using Nvidia GPUs provides between 10 and 14 times better price:performance than a conventional Azure HPC instance, said Hamilton.
Hybrid (Azure plus on premises) configurations are also supported, he said.
The announcement "really expands AI to the broadest set of organisations," said Hamilton.