When acquiring a supercomputer it is desirable to specify its performance using a single number. For many procurements, this is usually stated as a performance increase over a current generation platform, for example machine A provides 10 times greater performance than machine B. The determination of such a single number is not necessarily a simple process; there is no universal agreement on how this calculation is performed and each facility usually uses their own method. In the future, the landscape will be further complicated because systems will contain a heterogeneous mix of node types, and, by design, every application will not run on every node type. For example, at the National Energy Research Scientific Computing Center (NERSC) the Cori supercomputer contains two node types, nodes based on dual-socket Intel Xeon (Haswell) processors and nodes based on Intel Xeon Phi (Knights Landing) processors. However, NERSC evaluated these two partitions separately, without utilizing a single, combined performance metric. NERSC will be deploying its next-generation machine, NERSC-9, in the year 2020 and anticipates that it too will be a heterogeneous mix of node types. The purpose of this paper is to describe a single performance metric for a heterogeneous system.