Singularity: Microsoft to use public GPUs worldwide to build supercomputers

Microsoft has now for the first time publicly presented a system consisting of a large number of computers located around the world that can perform calculations with the utmost efficiency. The name of the infrastructure is Singularity. At the heart of Singularity is a workload balancing mechanism. This can analyze calculations in terms of their expected effort and urgency and then ensure that the operations are delivered to a global fleet of servers. Hundreds of thousands of GPUs and numerous FPGAs can also be used in Microsoft’s various data centers.

The main area of ​​application is, of course, training AI algorithms. This requires massive amounts of data to be evaluated on a regular basis. Normally this happens in a data center that needs to have some hardware on hand. However, a company the size of Microsoft and customers in the Azure cloud must constantly perform large amounts of AI calculations.

Decoupling of tasks and hardware

It, therefore, represents an enormous efficiency boost if the worldwide available resources can be used as optimally as possible. The tasks are decoupled from the hardware in the Singularity infrastructure, making it possible to flexibly assign physical devices to a project, depending on need and availability. Using a new method called replica splicing, multiple jobs can be processed simultaneously on a single server while minimizing the overhead required for control. “Singularity represents a major breakthrough in planning workloads for deep learning,” the statement said the paper, which was written by a number of Microsoft developers and describes the innovations developed over the course of Singularity’s development.