2018 OCP Regional Summit

Hardware specialization is the most effective path to higher performance and efficiency, which is attractive for large scale data centers.  Supporting many different hardware instances to get that specialization for different workloads, however, is expensive.  Each system must be separately qualified, different spares must be stocked, volumes are lower for each type of hardware increasing costs, and so on.  Microsoft achieves the benefits of specialization with common hardware by including an FPGA in every Azure and Bing server.  In this talk, I describe how Microsoft converged on a single FPGA platform that supports cutting edge performance on very different acceleration workloads, specifically accelerated software defined networking for Azure and deep learning inference for Bing.