Nearly seven months after they were first announced, Microsoft Azure’s HPC-targeted HBv2 virtual machines (VMs) based on AMD second-generation Epyc processors are ready for primetime. The new VMs, which Azure claims offer “supercomputer-class performance,” are already available in the South Central United States region and will soon be available in more U.S. regions, as well as in some of Europe and Japan.

The VMs are well-equipped: each one carries 120 AMD Epyc 7002-series CPU cores (clocking up to 3.3 GHz) and 480 GB of RAM with up to 340 GB/s of memory bandwidth. The HBv2s also offer MPI scalability up to 80,000 cores and are the first public VMs to offer 200 Gb/sec Mellanox HDR Infiniband, allowing latencies as low as 1.5 microseconds. Overall, Microsoft advertises a peak capability of four double-precision teraflops and eight single-precision teraflops per VM.

Regular pay-as-you-go pricing is $3.96 per hour, with spot instances starting as low as $0.901 per hour.

Microsoft advertises that the HBv2 VMs are suitable for HPC workloads ranging from computational fluid dynamics (CFD) and reservoir modeling to seismic processing and weather simulation. In the blog post that announced the new availability, Microsoft highlighted a couple of examples of how HBv2 VMs can be deployed in critical fields.

HBv2s in weather forecasting

If the past couple of months have demonstrated anything, it’s that numerical weather prediction (NWP) and supercomputing go hand-in-hand. Microsoft’s Cormac Garvey (a member of its HPC Azure Global team) used HBv2s to run the Weather Research and Forecasting (WRF) simulation suite as a demonstration of their ability in NWP. Garvey conducted a 371 million grid point, 1 kilometer resolution simulation of Hurricane Maria.

Garvey found that the HBv2s executed the simulation with “mostly superlinear scalability” up to 128 VMs, and that high scaling performance continued up to the largest scale (672 VMs). Microsoft says that the only reason scaling efficiency was limited in this case was that the model itself was too small for “such extreme levels of parallel processing.”

HBv2s in CFD for product design

Another member of the Azure Global HPC team, Jon Shelley, highlighted how HBv2 VMs could be used to run Simcenter Star-CCM+. To do this, Shelley worked with Siemens to validate a massive, billion-cell model of a sports car at a mesh resolution ten times higher than a similar test on Azure last year – one of the largest CFD simulations ever conducted with Simcenter Star-CCM+. In this simulation, Microsoft found linear efficiency up to 128 VMs and 84 percent scaling efficiency at the largest scale (640 VMs). 

To read more about about the launch, read the blog post by Evan Burness, principal program manager of Azure HPC, here. Burness also highlights the benefits of using Azure HBv2 VMs with parallel file systems BeeGFS and BeeGFS On Demand (BeeOND), a topic that is further explored in a blog post by Cormac Garvey.