This paper presents a novel system-level analysis of soft error rates (SER) based on the Transaction Level Model (TLM) of a targeted System-On-a-Chip (SoC). This analysis runs 1000x faster than the conventional SoC analysis using a gate-level model. Moreover, it allows accurate prediction in the early design phase of a SoC, when only limited application details are available. Preliminary validation results from accelerated SER tests on the physical system have shown that the analysis can predict the SER with a reasonable accuracy (within 5x of the results from tests on physical systems). This system-level analysis is particularly suitable to handle the black-box models for industrial semiconductor IP libraries. Based on this system-level analysis, we also propose a SE mitigation solution using selective protection of SRAM of a SoC. This solution provides a series of trade-offs between the system dependability and cost (in terms of silicon area).