Microsoft joined the Open Compute Project in 2014 and is a leading member. Last year we covered their contribution, Project Olympus, which was an Intel Skylake-EP platform. Microsoft has a challenge of needing to scale its cloud services, such as Azure, 10x every 4-5 years. That has led Microsoft to OCP and hardware design to meet its needs. Microsoft made a number of hardware and software announcements, integrating not only AMD, NVIDIA, and Intel but also Cavium and Qualcomm.
Microsoft at OCP
Universal PDU and universal motherboard. The Universal PDU has adapters for different electrical standards so that one PDU can be installed anywhere in the world.
Microsoft said that Project Olympus has a number of hardware partners.
Perhaps the most interesting here is the CPU box. Microsoft is welcoming AMD back into the datacenter with AMD Naples. It is also using Cavium and Qualcomm ARM64 chips. More on that soon.
Microsoft HGX-1 Hyperscale GPU Accelerator
The newly announced Microsoft HGX-1 is a hyperscale focused GPU machine for AI and machine learning. The Microsoft HGX-1 fits 8 GPUs per chassis using NVLINK.
You can also connect four chassis via PCIe which gives access to a total of 32GPUs in a system. Scaling outside of one system is a challenge in the AI community, so seeing Microsoft put together a large system like this is great.
Microsoft SONiC Open Source Software
Microsoft is highlighting its SONiC open source networking software that runs atop switches such as the OCP Wedge 100. Microsoft is not paying a Cisco premium and is opening up their networking software for others to use.
One of the cool features of SONiC is the ability to run containerized networking software and manage through Docker Swarm. This is the networking software behind Azure’s huge growth.
Optimizing ARM64 Servers for Microsoft Cloud
The biggest news from the OCP Summit is likely to be that Microsoft is heavily using ARM64 from both Cavium and Qualcomm. Qualcomm announced that Windows Server has been ported to its Centriq ARM server processors.
Microsoft said that it was a challenge to implement a new instruction set in the datacenter. We saw this while testing the Cavium ThunderX last year. Microsoft focused on porting Windows Server for Azure for internal purposes right now. Both Cavium and Qualcomm are running Windows server for Bing services at Microsoft. Here is a shot of the ARM motherboard:
Project Olympus building block architecture helped accelerate bringing ARM64 into Microsoft’s cloud services. Microsoft said without that, it would have been difficult to get the hardware design completed so quickly.