Floating vertical axis wind turbine (FVAWT) is a potential concept that could compete with the floating horizontal axis wind turbine (FHAWT) in the offshore wind industry. For a reliable FVAWT design, accurate prediction of extreme structural responses is crucial for ultimate limit state (ULS) design evaluation. Extreme structural responses of offshore wind turbines depend on operational modes and could present non-monotonic behavior with respect to environmental parameters. In this study, the extreme responses of a 3-bladed H-rotor semi-submersible FVAWT are evaluated using a simplified engineering approach, i.e., the modified environmental contour method (MECM). Judging from the response characteristics of the FVAWT with respect to increasing wind speed, the environmental contour with a return period of the cut-out wind speed is used. Along the selected environmental contour, the cut-out wind speed and the corresponding most probable sea state is chosen as the environmental design point. To generate statistical samples of structural responses, the Monte Carlo (MC) method is applied by carrying out fully-coupled time domain simulations under the selected environmental loadings. The Gumbel method, the Weibull tail method and the mean upcrossing rate (MUR) method are used to extrapolate the 50-yr long-term extreme responses. With a sufficiently large sample size, e.g. 100 MC simulations, the estimated extremes using different extrapolation methods are in close agreement with each other. However, the Weibull tail and MUR methods reduce the statistical uncertainty of the estimates due to higher data utilization. For practical engineering applications, massive amount of MC simulations is often not desirable. Therefore, the accuracy and statistical uncertainty of the estimated extreme responses associated with a reduced sample size is also investigated. For a sample of 10 simulations, the Weibull tail and MUR methods are able to predict extreme responses that are closer to the estimates using 100 simulations. The MUR method results in the lowest statistical uncertainty of the estimates.