Evaluation of RISC-V Silicon Under Neutron Radiation

Abstract

Radiation-hardened (rad-hard) components are frequently used in reliable spacecraft-computing systems. While these components improve mission dependability, they also suffer from high development and integration costs, support relatively low operating frequencies, and leverage outdated architectures. These characteristics motivate the consideration of more cost-effective and performant commercial alternatives. The open-source and highly configurable RISC-V architecture has recently become a popular choice for both space and commercial applications. While the reliability of RISC-V on FPGAs has been evaluated extensively, commercial RISC-V silicon has only begun to be investigated in a similar manner. This study evaluates the single-event upset (SEU) susceptibility of two commercial RISC-V processors, the Microchip PolarFire SoC and the SiFive HiFive Unmatched, in the presence of neutron radiation. These devices are compared to the flight-proven Xilinx Zynq-7020 system-on-chip, which contains an ARM Cortex-A9 processor. The industry-standard EEMBC CoreMark and SHREC-developed SpaceBench benchmarks are used to evaluate the presence of data and execution errors on each device under test (DUT). Neutron radiation beam testing was performed at the Los Alamos Neutron Science Center (LANSCE) Weapons Neutron Research (WNR) facility. Data- and execution-error results were recorded and analyzed to measure the proportion of errors present out of all calculations performed during the experiment. Effective dosimetry was also used to calculate cross sections of the processors that are susceptible to SEUs. The Po-larFire and Unmatched DUTs experienced no errors in 99.70% and 99.59% of operations, respectively. The Zynq achieved only 65.23% error-free operations. Execution errors were observed in 0.28%, 0.38%, and 18.67% of operations performed by the PolarFire, Unmatched, and Zynq, respectively. Similar trends were seen for data errors, with the PolarFire, Unmatched, and Zynq experiencing data errors in 0.02%, 0.03%, and 16.10% of operations, respectively. These results alongside dosimetry data produced cross sections of 8.033 × 10–12cm2 for the PolarFire and 8.342 × 10–12 cm2 for the Unmatched, indicating the area vulnerable to SEUs. The calculated cross section for the Cortex-A9 in the Zynq-7020 was 3.759 × 10–9 cm2a much larger value compared to either RISC-V platform. Both the error and cross-section analyses suggest that the evaluated commercial RISC-V devices have significantly lower SEU sus-ceptibility compared to the flight-proven Cortex-A9 platform, showing great promise for the reliable use of RISC-V silicon in embedded space applications.