Recent estimates suggest that the cpu speed of the fastest available microprocessors is increasing at approximately 80% per year [1], while the speed of memory devices has been growing at only about 7% per year [6]. As this trend continues, an increasing fraction of computer workloads will be dominated by memory access and transfer times rather than by compute time [2, 7]. This trend clearly suggests the adoption of memory latency and sustainable memory bandwidth as figures of merit for modern high-performance computers. Memory latency is measured by Larry McVoy's lmbench benchmark suite, while this report considers memory bandwidth.
The organization is as follows: First, the STREAM benchmark program is introduced, with an explanation of precisely what is being measured. Then a subset of the currently available results is presented, sorted by vendor and machine class. The interpretation of these results includes a review of some of the architectural and implementation factors which determine sustainable memory bandwidth. The paper concludes with a discussion of some trends observable in the data, and their implications for high performance computing.