Hybrid automatic repeat request (HARQ) is widely used in modern wireless communication systems to improve the spectral efficiency. However, the spectral efficiency improvement is at the cost of increased transmission delay. In this paper, we first show that the gap between spectral efficiency of conventional HARQ and ergodic channel capacity diminishes as $\frac {1}{D}$ , where $D$ is the average delay of HARQ. It is then shown that rate optimized HARQ, which allows static resource allocation of each HARQ transmission, is capable of achieving marginally better spectral efficiency and delay tradeoff performance. We then consider rate adaptive HARQ, which allows dynamic resource allocation based on outdated CSI feedback information from the receiver. Under the assumption of the cumulative distribution function of the fading channel capacity being a power function, spectral efficiency of simple linear stationary policy converges to the ergodic channel capacity at order of $\frac {1}{D^{2}}$ . For a wide class of policies, the spectral efficiency can converge to ergodic channel capacity at order of $e^{-D}$ . The optimal policy can be found in a simple explicit form sequentially by a greedy algorithm. Simulation results show that it achieves significantly higher spectral efficiency than other existing schemes with the same average delay.