Microsoft and Nvidia announced the new NDv2 at the SC19 Supercomputing Show in Denver. As a supercomputing solution based on GPU cloud acceleration, it will be adopted by Microsoft Azure. NDv2 virtual machines are known to be designed for demanding distributed high-performance computing, artificial intelligence, and machine learning workloads. These VMs come with eight NVLink-connected Nvidia Tesla V100 GPUs, complemented by a 32GB HBM2 cache, 40 non-hyperthreaded cores for the Intel Xeon Platinum 8168 processor, and 672 GB of system memory.
(Instagram via MSPU)
In addition, NDv2 has a 100 Gigabit EDR InfiniBand connection from Melanox that supports standard OFED drivers, all MPI types, and versions.
Microsoft says it is able to use these new virtual machines to train the popular session AI model BERT in just three hours.
Nvidia CEO Renxun Huang also said: Now you can open an instance, launch it in one of the Azure stacks (containers) and perform scientific calculations. This scheme makes it easy for every scientist in the world to have a supercomputer.