The very powerful turbo codes are a breakthrough in coding theory. The implementation of the iterative turbo decoder however hampers their incorporation in real systems. We tackle this problem by optimizing the decoder on two levels: the separate decoder modules and the global decoder system. Our module level optimizations greatly reduce the decoding delay and energy. The main contribution of this paper is situated at the system level, where we present an optimized architecture that uses a better interleaver organization. The combination of the optimizations on both levels results in a significant decoder power (factor 3), area (factor 2) and latency (factor 2) reduction.