Abstract
A novel compression technique for modulo-normalized state metrics in turbo decoder is presented. This technique performs the compression dynamically according to the range of state metrics. The framework of the compression circuit is given. BER performances for several turbo decoders are simulated. Implementation results show that the proposed technique needs only 25.2% logic gates and 77.5% memory bits of the existing method with same BER performance loss and much shorter critical path for eight state turbo codes. For four state turbo codes, 30.1% logic gates and 75% memory bits are required.
Introduction
Turbo codes are fundamental to modern high-speed wireless communication standards, such as 3GPP-LTE and WiMAX, due to their ability to approach the Shannon limit. However, the hardware implementation of turbo decoders is notoriously resource-intensive. A significant portion of this complexity arises from the need to store and process "state metrics" (SMs)—numerical values used in the Soft-Input Soft-Output (SISO) decoding process to track the likelihood of different bit sequences.
To prevent numerical overflow without sacrificing speed, hardware decoders often employ modulo-normalization. While effective, this technique requires substantial memory to store the full-precision metrics. Previous attempts to compress these metrics, such as using Walsh-Hadamard transforms, often introduced excessive hardware overhead or significant Bit Error Rate (BER) degradation. This paper proposes a dynamic compression technique that achieves high efficiency with minimal hardware cost and negligible performance loss.
The Challenge of Modulo-Normalized Metrics
In a modulo-normalized system, state metrics are represented on a numerical "circle." As the decoder progresses, these values can be scattered anywhere on this circle. Traditional static quantization fails here because it cannot account for the relative positions of the metrics.
The proposed method addresses this by observing that the absolute values of the metrics are less important than their relative differences. When the range of metrics is wide, the decoder is "confident" in its path, allowing for coarser quantization. Conversely, when the range is narrow, higher precision is maintained.
Proposed Dynamic Compression Framework
The researchers developed a four-step dynamic compression algorithm designed for simple hardware implementation:
- Rotation: To simplify the data, all state metrics are rotated so they cluster near zero. This is achieved by subtracting the value of the first state metric from all others in the set.
- Width Evaluation: The system determines the "Bit Width" ($BW$), which is the minimum number of bits required to represent the current range of the rotated metrics.
- Adaptive Quantization: The metrics are then compressed to a "Target Bit Width" ($TBW$). If the actual $BW$ exceeds the $TBW$, the circuit dynamically shifts the bits to fit the target width, effectively prioritizing the most significant bits.
- Restoration: Upon reading from memory, the process is reversed to reconstruct an approximation of the original metrics for the next stage of decoding.
Hardware Implementation
The architecture of the SISO decoder incorporating these units is shown below. The compression logic is integrated into the data path between the metric calculation units ($\alpha$, $\beta$, and $\gamma$ units) and the memory modules.
Unlike previous methods that required complex mathematical transforms, this "TM" (Target Metric) method relies on basic logic gates, OR-gates, and bit-shifters. This simplicity results in a significantly shorter critical path, enabling higher clock frequencies.
Experimental Results and Performance
The technique was evaluated using 3GPP-LTE (1024 bits), WiMAX (1920 bits), and generic 4-state turbo codes. Simulations were conducted over an AWGN channel using BPSK modulation with up to 8 iterations.
Key findings include:
- Logic Efficiency: For standard 8-state codes (LTE/WiMAX), the proposed method uses only 25.2% of the logic gates required by the state-of-the-art Walsh-Hadamard method.
- Memory Savings: It achieves a 22.5% reduction in memory bits (using 77.5% of the original requirement) while maintaining identical BER performance.
- Speed Improvements: The critical path was reduced from 3.2ns to 2.2ns, a 31% improvement that directly translates to higher throughput.
- 4-State Codes: For simpler decoders, the method saved 25% of memory and roughly 70% of logic gates.
Conclusion
The dynamic state metric compression technique presented here offers a highly efficient "middle path" for turbo decoder design. By leveraging the inherent confidence levels of the decoder to adjust quantization precision, the method slashes hardware area and memory requirements without the typical performance penalties. Future work may extend this framework to higher-radix decoders (e.g., radix-4) to meet the demands of next-generation ultra-high-speed wireless networks.
Summary updated