System Interconnect Architectures are Expected to Shift with Future HPC Procurements

Much attention is paid to the compute elements of HPC architectures, and rightly so. However, if those elements are not appropriately connected in a balanced and performant manner to other servers and to storage, then the optimizations provided by the server technologies will be unrealized as new bottlenecks appear. While there are multiple approaches for compute and storage interconnect architectures, recent survey data suggests the prevailing architecture is expected to shift from a preference of independent node-node and node-storage networks to a preference of converged networks with respondents’ next HPC procurement. This data is from an annual study which is part of the eighth edition of Hyperion Research’s high-performance computing (HPC) end-user-based tracking of the HPC marketplace. It included 181 HPC end-user sites with 3,830 HPC systems.