Abstract: We implement a 1-Tb/s 0.63-pJ/bit soft-assisted product decoder in a 28-nm technology. The decoder uses one bit of soft information to improve its net coding gain by 0.2 dB, reaching 10.3-10.4 dB, which is similar to that of more complex hard-decision staircase decoders.
Introduction
In high-speed fiber-optic communications, Forward Error Correction (FEC) is critical for maintaining signal integrity against noise. Designers typically face a trade-off between low-complexity Hard-Decision (HD) decoding and high-performance Soft-Decision (SD) decoding. While HD decoders are energy-efficient and suitable for Terabit speeds, they suffer from lower Net Coding Gain (NCG) because they discard signal confidence information. Conversely, SD decoders offer superior NCG but require complex fixed-point math and massive data storage, leading to high power consumption that is difficult to scale to Terabit speeds. This paper presents a "soft-assisted" product decoder that bridges this gap by utilizing a single bit of reliability information.
The iBDD-SR Architecture
The researchers implemented the iterative Bounded-Distance Decoding with Scaled Reliability (iBDD-SR) algorithm. This approach enhances standard HD decoding by incorporating just one extra bit of soft information per received bit to indicate reliability. This bit acts as a mask: if a bit is marked as "high-reliability," the decoder is prevented from flipping it, even if standard logic suggests an error. This effectively mitigates "miscorrections"—a common problem where HD decoders accidentally introduce new errors while trying to fix old ones.
The hardware implementation in 28-nm CMOS includes a dedicated reliability memory and minimal additional logic (one AND-gate per bit), resulting in only a 5% increase in circuit area and gate count compared to a pure HD decoder. The iterative process performs several rounds of row and column decoding, finding that a few soft-assisted rounds followed by standard HD rounds efficiently cleans up remaining errors.
Performance and Results
The decoder's performance was evaluated using a simulated Binary-Input Additive White Gaussian Noise (BI-AWGN) channel. Performance at the industry-standard $10^{-15}$ Bit Error Rate (BER) was estimated via extrapolation using the berfit tool in MATLAB. The results demonstrate that the single-bit soft assist provides a 0.2 dB NCG boost, reaching 10.3–10.4 dB. This performance is comparable to more complex staircase decoders but with lower latency and less memory.
The design achieves a peak throughput of 1 Terabit per second (Tb/s) with an energy efficiency of 0.63 pJ/bit at 5 iterations. As shown in the table below, there is a clear trade-off between the number of iterations, throughput, and energy consumption. For instance, increasing to 10 iterations improves error correction but reduces throughput to 516 Gb/s and increases energy cost to 0.97 pJ/bit.
Conclusion and Future Work
This work demonstrates that high-performance error correction does not necessarily require high-precision soft information. By using just one bit of reliability, the proposed decoder achieves state-of-the-art NCG while maintaining the extreme energy efficiency and throughput required for next-generation fiber-optic networks. While the current design uses a fixed scaling parameter for reliability to maintain simplicity, future work may involve dynamic adjustment based on channel conditions to potentially yield even higher gains.
Summary updated