Various volatility estimators and models have been proposed in the literature to measure volatility of asset returns. In this paper, we compare empirical performance of various unconditional volatility estimators and conditional volatility models (GARCH and EGARCH) using time-series data of S&PCNX Nifty, a value-weighted index of 50 stocks traded on the National Stock Exchange (NSE), Mumbai. The estimates computed by various estimators and conditional volatility models over non-overlapping one-day, five-day and one-month periods are compared with the "realized volatility" measured over the same period. We use three years' (1999-2001) high-frequency data set of five-minute returns to construct measures of realized volatility. In order to test the ability of the estimators and models to forecast volatility, we compare the estimates of unconditional estimators with the realized volatility measured in the next period of same length. For conditional volatility models, the forecasts for the same periods are obtained by estimating models from the time-series prior to the forecast period. Our results indicate that while conditional volatility models provide less biased estimates, extreme-value estimators are more efficient estimators of realized volatility. As far as forecasting ability of models and estimators is concerned, conditional volatility models fare extremely poorly in forecasting five-day (weekly) or monthly realized volatility. In contrast, extreme-value estimators, other than the Parkinson estimator, perform relatively well in forecasting volatility over these horizons.