By: Michael Feldman

Microsoft is introducing yet another set of high-performance Azure instances for its HPC and AI customers.

The new instances were part of a larger group of announcements outlining a wide array of Azure hardware and software enhancements. Like most public clouds, Azure is undergoing near-constant upgrades, which get rolled out, or at least teased, at the company’s next available event. In the case of the HPC and AI upgrades, Microsoft used its Ignite conference in Orlando, Florida to reveal new Azure infrastructure powered by NVIDIA Tesla GPUs, Intel Xeon CPUs, and AMD EPYC processors.

The additional GPU instance, known as NVv2, is based on M60-powered servers that are primarily targeted at remote visualization and other graphics-demanding workloads. As is implied by its v2 designation, Azure already offers an M60-powered instance– the GPU itself was introduced in 2015 – but this new configuration doubles up on system memory capacity (up to 448 GiB) and adds support for “Premium SSDs.” According to Microsoft, the extra memory and SSD support will make the NVv2 optimal for things like CAD, gaming and interactive 3D design. The NVv2 instances are available in preview now in the West US, South Central US, and West Europe regions.

Microsoft is also planning to add NVIDIA V100 GPU instances to its ND series, which are aimed at training and inferencing deep learning models. The current ND configurations are based on the older P40 GPUs. (Azure currently has other V100 instances in its NC series, but these are targeted to conventional HPC workloads.) The new NDv2 instances will be powered by up to eight NVIDIA Tesla V100 GPUs, 40 Intel Skylake cores, and 672 GiB of main memory. These instances are slated to be in preview by the end of the year.

Microsoft is also planning to roll out new HB and HC series instances, both of which are designed for traditional high performance computing. HB and HC are offshoots of the existing H series instances, which are currently powered by the older Haswell Xeon CPUs. Basically, the HB series will be offered for HPC applications driven by high memory bandwidth requirements, while the HC series will be targeted to more compute-intensive workloads.

As a result, the new HB instance (HB60rs) will use the EPYC CPUs, which offer about 33 more memory bandwidth than their Xeon competition. For applications that employ things like fluid dynamics, explicit finite element analysis, and weather modeling, the higher bandwidth offers a distinct advantage over the more flops-capable Xeons. The HB instance will offer 60 EPYC 7551 processor cores, with 4 GB of RAM per CPU core.

This is not Azure’s first go-around with EPYC silicon. In 2017, AMD announced Microsoft would use its processors in Azure’s L series instances for storage-demanding workloads. However, all the L series instances currently on the Azure price list specify Xeon processors. Hopefully, the EPYC servers in the HB series will not suffer the same fate.

As you may have already surmised, the new HC series will employ Intel’s Xeon Scalable processors, which, thanks to their AVX-512 capability, offer significantly more floating point performance than the EPYC chips. The new HC44rs instance will provide 44 Xeon Platinum 8168 cores, with 8 GB of RAM per core. It is geared for flops-demanding codes, such as implicit finite element analysis, reservoir simulation, and computational chemistry.

The new EPYC instance will be available for preview later this year in the South Central US region, while the new Xeon instance will be available in the same time frame in the West US 2 region. Microsoft is planning to roll them out across additional regions soon thereafter.