Non-binary low-density parity-check (NB-LDPC) codes can achieve better error-correcting performance than binary LDPC codes when the code length is moderate at the cost of higher decoding complexity. The high complexity is mainly caused by the complicated computations in the check node processing and the large memory requirement. In this thesis, two VLSI designs for NB-LDPC decoders based on two novel check node processing schemes are proposed. The first design is based on forward-backward check node processing. A novel scheme and corresponding architecture are developed to implement the elementary step of the check node processing. In our design, layered decoding is applied and only nm less than q messages are kept on each edge of the associated Tanner graph. The computation units and the scheduling of the computations are optimized in the context of layered decoding to reduce the area requirement and increase the speed. This thesis also introduces an overlapped method for the check node processing among different layers to further speed up the decoding. From complexity and latency analysis, our design is much more efficient than any previous design. Our proposed decoder for a (744, 653) code over GF(32) has also been synthesized on a Xilinx Virtex-2 Pro FPGA device. It can achieve a throughput of 9.30 Mbps when 15 decoding iterations are carried out. The second design is based on a proposed trellis based check node processing scheme. The proposed scheme first sorts out a limited number of the most reliable variable-to-check (v-to-c) messages, then the check-to-variable (c-to-v) messages to all connected variable nodes are derived independently from the sorted messages without noticeable performance loss. Compared to the previous iterative forward-backward check node processing, the proposed scheme not only significantly reduced the computation complexity, but eliminated the memory required for storing the intermediate messages generated from the forward and backward processes. Inspired by this novel c-to-v message computation method, we propose to store the most reliable v-to-c messages as 'compressed' c-to-v messages. The c-to-v messages will be recovered from the compressed format when needed. Accordingly, the memory requirement of the overall decoder can be substantially reduced. Compared to the previous Min-max decoder architecture, the proposed design for a (837, 726) code over GF(32) can achieve the same throughput with only 46% of the area.

Non-binary low-density parity-check (NB-LDPC) codes can achieve better error-correcting performance than their binary counterparts when the code length is moderate at the cost of higher decoding complexity. The high complexity is mainly caused by the complicated computations in the check node processing and the large memory requirement. In this thesis, three decoding algorithms and corresponding VLSI architectures are proposed for NB-LDPC codes to lower the computational complexity and memory requirement. The first design is based on the proposed relaxed Min-max decoding algorithm. A novel relaxed check node processing scheme is proposed for the Min-max NB-LDPC decoding algorithm. Each finite field element of GF(2p̂) can be uniquely represented by a linear combination of $p$ independent field elements. Making use of this property, an innovative method is developed in this paper to first find a set of the p most reliable variable-to-check messages with independent field elements, called the minimum basis. Then the check-to-variable messages are efficiently computed from the minimum basis. With very small performance loss, the complexity of the check node processing can be substantially reduced using the proposed scheme. In addition, efficient VLSI architectures are developed to implement the proposed check node processing and overall NB-LDPC decoder. Compared to the most efficient prior design, the proposed decoder for a (837, 726) NB-LDPC code over GF(25̂) can achieve 52% higher efficiency in terms of throughput-over-area ratio. The second design is based on a proposed enhanced iterative hard reliability-based majority-logic decoding. The recently developed iterative hard reliability-based majority-logic NB-LDPC decoding has better performance-complexity tradeoffs than previous algorithms. Novel schemes are proposed for the iterative hard reliability-based majority-logic decoding (IHRB-MLGD). Compared to the IHRB algorithm, our enhanced (E- )IHRB algorithm can achieve significant coding gain with small hardware overhead. Then low-complexity partial-parallel NB-LDPC decoder architectures are developed based on these two algorithms. Many existing NB-LDPC code construction methods lead to quasi-cyclic or cyclic codes. Both types of codes are considered in our design. Moreover, novel schemes are developed to keep a small proportion of messages in order to reduce the memory requirement without causing noticeable performance loss. In addition, a shift-message structure is proposed by using memories concatenated with variable node units to enable efficient partial-parallel decoding for cyclic NB-LDPC codes. Compared to previous designs based on the Min-max decoding algorithm, our proposed decoders have at least tens of times lower complexity with moderate coding gain loss. The third design is based on a proposed check node decoding scheme using power representation of finite field elements. Novel schemes are proposed for the Min-max check node processing by making use of the cyclical-shift property of the power representation of finite field elements. Compared to previous designs based on the Min-max algorithm with forward-backward scheme, the proposed check node units (CNUs) do not need the complex switching network. Moreover, the multiplications of the parity check matrix entries are efficiently incorporated. For a Min-max NB-LDPC decoder over GF(32), the proposed scheme reduces the CNU area by at least 32%, and leads to higher clock frequency.

The VLSI implementation complexity of a low density parity check (LDPC) decoder is largely influenced by the interconnect and the storage requirements. This dissertation presents the decoder architectures for regular and irregular LDPC codes that provide substantial gains over existing academic and commercial implementations. Several structured properties of LDPC codes and decoding algorithms are observed and are used to construct hardware implementation with reduced processing complexity. The proposed architectures utilize an on-the-fly computation paradigm which permits scheduling of the computations in a way that the memory requirements and re-computations are reduced. Using this paradigm, the run-time configurable and multi-rate VLSI architectures for the rate compatible array LDPC codes and irregular block LDPC codes are designed. Rate compatible array codes are considered for DSL applications. Irregular block LDPC codes are proposed for IEEE 802.16e, IEEE 802.11n, and IEEE 802.20. When compared with a recent implementation of an 802.11n LDPC decoder, the proposed decoder reduces the logic complexity by 6.45x and memory complexity by 2x for a given data throughput. When compared to the latest reported multi-rate decoders, this decoder design has an area efficiency of around 5.5x and energy efficiency of 2.6x for a given data throughput. The numbers are normalized for a 180nm CMOS process. Properly designed array codes have low error floors and meet the requirements of magnetic channel and other applications which need several Gbps of data throughput. A high throughput and fixed code architecture for array LDPC codes has been designed. No modification to the code is performed as this can result in high error floors. This parallel decoder architecture has no routing congestion and is scalable for longer block lengths. When compared to the latest fixed code parallel decoders in the literature, this design has an area efficiency of around 36x and an energy efficiency of 3x for a given data throughput. Again, the numbers are normalized for a 180nm CMOS process. In summary, the design and analysis details of the proposed architectures are described in this dissertation. The results from the extensive simulation and VHDL verification on FPGA and ASIC design platforms are also presented.

Near-capacity performance and parallelizable decoding algorithms have made Low-Density Parity Check (LDPC) codes a powerful competitor to previous generations of codes, such as Turbo and Reed Solomon codes, for reliable high-speed digital communications. As a result, they have been adopted in several emerging standards. This thesis investigates VLSI architectures for multi-Gbps power and area-efficient LDPC decoders. To reduce the node-to-node communication complexity, a decoding scheme is proposed in which messages are transferred and computed bit-serially. Also, a broadcasting scheme is proposed in which the traditional computations required in the sum-product and min-sum decoding algorithms are repartitioned between the check and variable node units. To increase decoding throughput, a block interlacing scheme is investigated which is particularly advantageous in fully-parallel LDPC decoders. To increase decoder energy efficiency, an efficient early termination scheme is proposed. In addition, an analysis is given of how increased hardware parallelism coupled with a reduced supply voltage is a particularly effective approach to reduce the power consumption of LDPC decoders. These architectures and circuits are demonstrated in two hardware implementations. Specifically, a 610-Mbps bit-serial fully-parallel (480, 355) LDPC decoder on a single Altera Stratix EP1S80 device is presented. To our knowledge, this is the fastest FPGA-based LDPC decoder reported in the literature. A fabricated 0.13-mum CMOS bit-serial (660, 484) LDPC decoder is also presented. The decoder has a 300 MHz maximum clock frequency and a 3.3 Gbps throughput with a nominal 1.2-V supply and performs within 3 dB of the Shannon limit at a BER of 10-5. With more than 60% power saving gained by early termination, the decoder consumes 10.4 pJ/bit/iteration at Eb=N0=4dB. Coupling early termination with supply voltage scaling results in an even lower energy consumption of 2.7 pJ/bit/iteration with 648 Mbps decoding throughput. The proposed techniques demonstrate that the bit-serial fully-parallel architecture is preferred to memory-based partially-parallel architectures, both in terms of throughput and energy efficiency, for applications such as 10GBase-T which use medium-size LDPC code (e.g., 2048 bit) and require multi-Gbps decoding throughput.

Since the rediscovery of low-density parity-check (LDPC) codes in the late 1990s, tremendous progress has been made in code construction and design, decoding algorithms, and decoder implementation of these capacity-approaching codes. Recently, LDPC codes are considered for applications such as high-speed satellite and optical communications, the hard disk drives, and high-density flash memory based storage systems, which require that the codes are free of error-floor down to bit error rate (BER) as low as 10−12 to 10−15. FPGAs are usually used to evaluate the error performance of codes, since one can exploit the finite word length and extremely high internal memory bandwidth of an FPGA. Existing FPGA-based LDPC decoders fail to utilize the configurability and read-first mode of embedded memory in the FPGAs, and thus result in limited throughput and codes sizes. Four optimization techniques, i.e., vectorization, folding, message relocation, and circulant permutation matrix (CPM) sharing, are proposed to improve the throughput, scalability, and efficiency of FPGA-based decoders. Also, a semi-automatic CAD tool called QCSYN (Quasi-Cyclic LDPC decoder SYNthesis) is designed to shorten the implementation time of decoders. Using the above techniques, a high-rate (16129,15372) code is shown to have no error-floor down to the BER of 10−14. Also, it is very difficult to construct codes that do not exhibit an error floor down to 10−15 or so. Without detailed knowledge of dominant trapping sets, a backtracking-based reconfigurable decoder is designed to lower the error floor of a family of structurally compatible quasi-cyclic LDPC codes by one to two orders of magnitudes. Hardware reconfigurability is another significant feature of LDPC decoders. A tri-mode decoder for the (4095,3367) Euclidean geometry code is designed to work with three compatible binary message passing decoding algorithms. Note that this code contains 262080 edges (21.3 times of the (2048,1723) 10GBASE-T code) in its Tanner graph and is the largest code ever implemented. Besides, an efficient QC-LDPC Shift Network (QSN) is proposed to reduce the interconnect delay and control logic of circular shift network, a core component in the reconfigurable decoder that supports a family of structurally compatible codes. The interconnect delay and control logic area are reduced by a factor of 2.12 and 8, respectively. Non-binary LDPC codes are effective in combating burst errors. Using the power representation of the elements in the Galois field to organize both intrinsic and extrinsic messages, we present an efficient decoder architecture for non-binary QC-LDPC codes. The proposed decoder is reconfigurable and can be used to decode any code of a given field size. The decoder supports both regular and irregular non-binary QC-LDPC codes. Using a practical metric of throughput per unit area, the proposed implementation outperforms the best implementations published in research literature to date.

Binary Low-Density Parity-Check (LDPC) codes are a type of error correction code known to exhibit excellent error-correcting capabilities, and have increasingly been applied as the forward error correction solution in a multitude of systems and standards, such as wireless communications, wireline communications, and data storage systems. In the pursuit of codes with even higher coding gain, non-binary LDPC (NB-LDPC) codes defined over a Galois field of order q have risen as a strong replacement candidate. For codes defined with similar rate and length, NB-LDPC codes exhibit a significant coding gain improvement relative to that of their binary counterparts. Unfortunately, NB-LDPC codes are currently limited from practical application by the immense complexity of their decoding algorithms, because the improved error-rate performance of higher field orders comes at the cost of increasing decoding algorithm complexity. Currently available ASIC implementation solutions for NB-LDPC code decoders are simultaneously low in throughput and power-hungry, leading to a low energy efficiency. We propose several techniques at the algorithm level as well as hardware architecture level in an attempt to bring NB-LDPC codes closer to practical deployment. On the algorithm side, we propose several algorithmic modifications and analyze the corresponding hardware cost alleviation as well as impact on coding gain. We also study the quantization scheme for NB-LDPC decoders, again in the context of both the hardware and coding gain impacts, and we propose a technique that enables a good tradeoff in this space. On the hardware side, we develop a FPGA-based NB-LDPC decoder platform for architecture prototyping as well as hardware acceleration of code evaluation via error rate simulations. We also discuss the architectural techniques and innovations corresponding to our proposed algorithm for optimization of the implementation. Finally, a proof-of-concept ASIC chip is realized that integrates many of the proposed techniques. We are able to achieve a 3.7x improvement in the information throughput and 23.8x improvement in the energy efficiency over prior state-of-the-art, without sacrificing the strong error correcting capabilities of the NB-LDPC code.