GPU-Accelerated Optimal Interpolation for Global Ocean Surface pCO2 Mapping

GPU acceleration has become essential for meeting the rising computational demands of high-resolution ocean modeling and data assimilation. In this study, a GPU-accelerated, chunk-based, sequential Optimal Interpolation (OI) scheme was developed to reconstruct global ocean surface partial pressure of CO2 (pCO2) fields. OI analyses were successfully performed for the period between 1990 and 2019 using 30 years of monthly background fields and observations. Benchmarking highlighted substantial gains, with runtimes up to 35 times faster than that of a single-core CPU baseline and approximately 12 times faster than that of multi-core CPU runs. The GPU performance improved steadily with increasing chunk size, whereas the fastest runtimes were consistent with observation batch sizes in the range of 1,000–2,000. Sequential OI generated analysis fields nearly identical to those from the all-at-once observation update, and its runtime ranged from a bit slower to slightly faster depending on the choice of observation batch size. These results show that GPU-based OI is both practical and efficient, offering a pathway for the direct application of GPUs in operational ocean prediction systems.