Today at its AzureCon, Microsoft expanded the capabilities of its public cloud, Azure, with the addition of N-series GPU-enabled virtual machines available over a fast RDMA network. The company also announced that it is reducing prices on its high-end instances, A8-A11.
Jazon Zander, corporate vice president at Microsoft Azure, began by presenting a view of Azure instance offerings arranged from the highest value to largest scale-up:
In addition to Azure’s A, D and G series of virtual machine types, Zander announced two new options. The first, called DV2, is based on the latest generation 2.4 GHz Intel Xeon E5-2673 v3 (Haswell) processor. DV2 series instances are on average 35 percent faster than the D series and are intended for applications that demand faster CPUs, better local disk performance, or higher memories. With the announcement of DV2, Microsoft said it is lowering D-series prices across all regions.
The second new instance family, the GPU-backed N-series, provides Azure customers access to the NVIDIA Tesla K80 GPU accelerator for HPC workloads and to NVIDIA GRID 2.0 technology for virtualized graphics in enterprise workflows. The low-latency Remote Direct Memory Access (RDMA) network will facilitate remote visualization, high-performance computing and analytics workloads, said Zander.
NVIDIA CEO Jen-Hsun Huang spoke over video about the collaboration to make accelerated graphics and high performance computing available over any connected device.
“It takes an enormous amount of innovation in order to put GPUs into the cloud,” said Jen-Hsun Huang. “These are processors that are intensely high-performance but want to be very close to the user. Through very important innovations, we’ve now made it possible for these powerful GPUs to be included in the cloud. The first is the intense energy-efficiency focus that is required because datacenters are really large and the cost to manage the datacenter is very high. The second element is the shortest possible latency so that the computer graphics that is generated or the computation that is done is instantly transmitted to the user. And the third: a GPU has to be general purpose. One of the most important characteristics of the NVIDIA GPUs [comes from] the two decade long effort that we put into compatibility and the performance optimizations.”
“For very first time,” he continued, “we have virtualized the entire software stack and integrated it deeply into the Microsoft Azure cloud. As a result, we can be completely compatible with almost any application you can think of. It could be a seismic processing application, a car design application, an industrial design application, a product design application to medical imaging applications. There are no applications that run on a PC today that are not compatible with the NVIDIA graphics stack, and now that capability is put into the Azure cloud.”
The Azure public cloud is the first to offer NVIDIA GRID 2.0 virtualized graphics, powered by the Tesla M60 GPU. Instances backed with the NVIDIA K80 GPU will support HPC, computational science, data analytics and deep learning applications. Full RDMA provides low-latency network connection for multi-machine GPU workloads.
Zander said that Microsoft is working with a number of ISVs that specialize in computer-aided design and simulation to bring more of their workloads onto Azure.
The N-series instances will be available in preview within the next few months. Prices and configurations have not been disclosed.
Azure’s other HPC-friendly instances, A8-A11, haven’t had the Haswell-refresh yet, but the company did say it was dropping prices by as much as 60 percent, effective October 1, 2015. The rate reductions apply to the network-optimized A8 and A9, which feature a 40Gbit/s InfiniBand network with RDMA technology, and the compute-optimized A10 and A11, which are identical to A8 and A9, except without the InfiniBand network and RDMA technology.
“These instances carry the powerful Intel Xeon E5 processors and are suitable for compute intensive workloads like high-performance clusters, modeling and simulations, video encoding, and other compute or network intensive scenarios,” wrote Zander in a blog post, directing interested users to the product pricing page.