The proposed LDPC decoder has a partially parallel
architecture with 16 BNUs, 16 CNUs and exchange messages
with 16 shared memory modules. Each CNU reads the
memory from the address generated by the AGU and
processes the operation. After completing each operation, the
message to be delivered is stored in the same address. It is
crucial for the AGU to generate precise addresses without
causing any memory access conflict. Therefore, we propose a
novel approach which enables us to parallelize processing
with 16 shared memories without any conflict. Also, we
propose a highly efficient AGU design which generates
addresses based on an Index matrix. Furthermore, to support
two data rates with a minimal area overhead, most of the
blocks are reconfigurable so that they can be commonly used
for both modes.