Efficient simulations of quantum evolutions of spin-1/2 systems are relevant for ensemble quantum computation as well as in typical NMR experiments. We propose an efficient method to calculate the dynamics of an observable provided that the initial excitation is "local". It resorts a single entangled pure initial state built as a superposition, with random phases, of the pure elements that compose the mixture. This ensures self-averaging of any observable, drastically reducing the calculation time. The procedure is tested for two representative systems: a spin star (cluster with random long range interactions) and a spin ladder.PACS numbers: 03.67. Lx, 05.30.Ch, 75.40.Gb, 75.10.Pq One of the goals of quantum computers is the simulation of quantum systems. While quantum processing is ideally cast in terms of pure states, experimental realizations are a major challenge [1] met only for very small systems [2]. The alternative use of statistical mixtures of pure states, as the spin ensembles in standard NMR experiments [3,4,5,6], led to the development of the ensemble quantum computation (EQC) [7] that allowed useful algorithm optimizations [8,9]. Recently, EQC was experimentally implemented in a 12-qubits system [10] and much larger quantum registers [11] were prepared to assess its stability against decoherence. Hence, an efficient evaluation of ensemble evolutions is needed. Analytical solutions of ensemble dynamics, either from the integration of the Liouville von-Neumann equation [12] or the alternative Keldysh formalism [13,14] are limited to special cases (e.g. [15,16]). Thus, one has to resort to numerical solutions. One can describe an ensemble of M spins by using the 2 M × 2 M density matrix, but this quickly reaches a storage limit as M increases. Thus, a typical exact diagonalization method in a desktop computer slightly exceeds a dozen of qubits. Alternatively, the use of wave functions combined with the TrotterSuzuki decomposition [17,18] overcomes this limitation because it uses vectors of size 2 M . However, the average over individual evolutions of a large number of components of the ensemble takes a long time. Here, this limitation is overcome by profiting of quantum parallelism [19] to evaluate the ensemble dynamics of any observable evolved from a "local" initial condition. The idea is that when evaluated on single pure states that are a superposition of all the elements of the ensemble, these observables become self-averaging. This reinforces the suggestions that one can avoid ensemble or thermal averages by using a single pure state [20,21]. Ref.[20] considers a subsystem with the reduced density matrix derived from a pure state, where the subsystem is entangled with an environment which has a much bigger size. The resulting reduced density matrix describes the microcanonical ensemble without resorting to the equiprobability postulate of statistical mechanics. Following a similar inspiration, we focus on the non-equilibrium dynamics of any given observable. The key is that the initial non-eq...