The paper proposes a random linear model to investigate the memory bandwidth barrier against performance of sparse iterative methods on shared memory computers. By comparing the performance of sparse matrix vector multiplications in different formats, the paper shows that one of the most important kernels for sparse iterative methods is memory intensive operation and its realistic performance in the traditional floating-point operations metric is limited by the memory bandwidth. Therefore, a fair metric for performance of sparse iterative method should consider the memory bandwidth capability and memory efficiency. Various numerical results are presented, compared, analysed and validated to confirm the proposed model.