In this paper, we propose a semiparametric Bayesian model for extreme value analysis with big spatiotemporal data. We consider a Dirichlet process mixture (DPM) of Gaussian processes with random location-scale components. We consider different choices for the probability distributions of the random locations and scales of the mixing components and discuss the bulk and the tail properties. In our model, the temporal replications are clustered automatically and based on the clusters of extremes, we aim at jointly identifying the hot and cold spots. Thus, our approach is free from truncating the data above/below some arbitrary thresholds or considering only the location-wise maximums and minimums. Considering very high spatial dimension as well as nonstationarity, we allow low-rank approximation of the Gaussian process components using empirical orthogonal functions in order to allow fast computation. Inference is drawn based on Markov chain Monte Carlo sampling. We use the proposed model to identify the hot and cold spots for the sea surface temperature of the Red sea.