Scanning the Issue

Traditionally, cloud platforms have been based on a single computing device type: central processing units (CPUs). One of the main reasons for this homogeneity of hardware resources has been cost efficiency—for years, cloud providers have reaped the benefits of the economies of scale by buying thousands of very similar types of servers. The homogeneity of servers has other advantages as well, for instance, easy management and scheduling of resources, and simple development and deployment of applications and tools for debugging and tracing. However, in recent years, cloud servers have undergone a significant change. They have progressively shifted to become heterogeneous platforms in which CPUs join forces with special-purpose integrated circuits, e.g., Google’s tensor processing units (TPUs), graphics processing units (GPUs), and field-programmable gate arrays (FPGAs).