Microsoft reveals its new Azure HPC offerings at SC19
At SC19 in Denver, Microsoft demonstrated its new HPC Azure hardware based on NVIDIA and Graphcore accelerators.
The new Microsoft Azure Virtual Machines are part of the Open Compute Project (OCP), a leading-edge open hardware development model that promotes possibilities for global data center deployment in the cloud. The platform is called Open CloudServer (previously, Project Olympus). All design specifications are open-source and standardized, which allows IT and data center operators to adjust community-developed innovation and scale proven hardware designs for their organizations’ specific purposes.
The new product line opens up with Azure NDv2 VMs, the fastest and the most powerful among the rest of Azure’s offerings. They feature 8 NVIDIA Tesla V100 NVLink interconnected GPUs each with 32 GB of HBM2 memory, 40 non-hyperthreaded cores from the Intel Xeon Platinum 8168 processor, and 672 GiB of system memory. NDv2 VMs are specifically designed for the most demanding machine learning and AI workloads.
HBv2-series VMs are positioned as one of the most affordable and yet the most powerful HPC servers Azure has ever made. They feature 120 non-hyperthreaded CPU cores from the new AMD EPYC™ 7742 processor, up to 45 percent more memory bandwidth than x86 alternatives, and up to 4 teraFLOPS of double-precision compute.
With 200 Gigabit HDR InfiniBand from Mellanox and an ability to support up to 80,000 cores per MPI job, HBv2 series VMs can deliver workload scalability and performance comparable to the world’s most powerful bare metal supercomputers.
Azure NDv3 VMs are the first series in the product line to feature the unique Graphcore IPU architecture, specifically designed to accelerate AI training and inference workloads via high-throughput processing of neural networks even at small batch sizes.
NDv3 VMs feature 16 IPU chips, each with over 1200 cores and over 7000 independent threads and a large 300MB on-chip memory that delivers up to 45 TB/s of memory bandwidth. The NDv3 series’ interconnect system enables a model-parallel or data parallel training for large neural networks models.
Finally, Azure NVv4 VMs have been presented as the most flexible solution, offering customers partionable GPUs. Thus, customers can select a virtual GPU size that matches their organization’s specific needs and price points.
NVv4 VMs make a good addition to Azure’s portfolio of Windows Virtual Desktops, featuring the new AMD EPYC™ 7742 processor and the AMD RADEON INSTINCT™ MI25 GPUs.
All of Azure’s new offerings align with the company’s goal to develop HPC-oriented cloud solutions that can either rival or exceed some of the most sought-after on-premises deployments.