Terrestrial-based radiation upsets: A cautionary tale

Problems with terrestrial-based neutron radiation from cosmic rays have become more commonplace. While the incident rate from neutron radiation is lower than space-based radiation, physics, system design and system locations have combined to make systems in-creasingly vulnerable to terrestrial radiation. FPGA systems are particularly sensitive to neutron radiation, as the FPGAs, microprocessors and memory are all sensitive to upsets. We are interested in reconfig-urable supercomputers,which need to be highly reliable and highly available despite being very sensitive to ra-diation. In this paper, we estimate the error rate for FPGAs, memory, and microprocessors so that predic-tions for the sensitivity of the Cray XD1 reconfigurable supercomputer can be made. We also present possible mitigation methods that are appropriate for neutron ra-diation upset rates.