On fault tolerant decoding of Turbo codes

Decoding of Turbo Codes requires buffer memories to store the received values and the extrinsic information that is exchanged between the constituent decoders. In this paper, the effect of unreliable buffer memories on the decoding performance is analyzed. The buffer is modeled as a discrete memoryless channel, which introduces spatially independent and uniform bit errors on the binary representation of the stored values. This leads to a strong performance degradation if a conventional Turbo decoding algorithm is employed. It is however shown that suitable modification of quantizer, index assignment, and of the transition metrics of the MAP algorithm can effectively compensate for these errors.