Introduction
In the harsh radiation environments of space, semiconductor devices face constant bombardment from high-energy particles. These interactions lead to Single-Event Upsets (SEUs) and Single-Event Transients (SETs), which can corrupt data or disrupt control logic. While significant research has focused on hardening CPU cores and memory subsystems, the on-chip interconnect—the "glue" of the System-on-Chip (SoC)—remains a critical single point of failure. A single bit flip in a request signal or address bus can lead to system hangs, incorrect memory writes, or infinite loops, necessitating a full system reset.
The Challenge of Interconnect Reliability
Traditional reliability solutions for Networks-on-Chip (NoCs), such as retransmission protocols, introduce significant latency and are often unsuitable for tightly-coupled, low-latency interconnects found in microcontrollers. Conversely, "brute force" hardware solutions like fine-grained Triple Modular Redundancy (TMR) provide high reliability but at a prohibitive cost, often increasing the silicon area by 4x to 7x. This overhead is frequently unacceptable for power- and cost-constrained space missions.
The relOBI Architecture
Researchers from ETH Zurich and the University of Bologna have proposed relOBI, a reliable extension to the Open Bus Interface (OBI) protocol. relOBI employs a hybrid protection strategy that prioritizes signals based on their functional criticality to achieve high reliability with optimized area overhead.
Hybrid Protection Strategy
The core innovation of relOBI lies in its protocol-aware application of redundancy:
- TMR for Handshake Signals: Critical control signals, such as request (
req), grant (gnt), and read valid (rvalid), are triplicated and passed through majority voters. This ensures that the protocol's state machine remains synchronized even in the presence of single-bit faults. - ECC for Payload Signals: Data buses, addresses, and byte enables are protected using Error Correction Codes (ECC). This approach is significantly more area-efficient than triplicating wide buses (e.g., 32-bit or 64-bit).
Hardened Crossbar and Timing Optimization
The relOBI implementation includes a hardened 6×8 crossbar. To mitigate the latency penalties associated with ECC decoding, the authors introduced an "optimized" path. This design checks for errors in parallel with the request processing. If an error is detected, the transaction is aborted and corrected in the subsequent cycle, maintaining low latency for the common (error-free) case.
Experimental Evaluation
The design was evaluated through extensive fault injection campaigns using Synopsys VC Z01X. The researchers injected 1,000,000 random faults into three distinct designs: a baseline (unprotected) OBI crossbar, the proposed relOBI crossbar, and a fully triplicated crossbar (TMRG). Faults were injected at both the RTL level (registers and ports) and the Netlist level (logic primitives).
Key Results
The experimental results demonstrate that relOBI successfully eliminates vulnerabilities to single-bit faults while maintaining superior area efficiency compared to full triplication.
- Reliability: In simulations involving 1 million faults, relOBI reduced the undetected error rate from 34.85% in the baseline design to zero.
- Area Efficiency: relOBI required a 2.6× area increase over the baseline. While substantial, this is 1.8× lower than the 4.7× increase required by the fully triplicated TMRG design.
- Bus Overhead: Converting a standard 137-bit OBI bus to relOBI results in a 177-bit bus (a 29% increase), compared to the 200% increase required for full triplication.
- Performance: The optimized relOBI design achieved a timing of 19.1 FO4 (fan-out-of-4 delay). Although slower than the baseline (13.6 FO4), it remains capable of supporting clock frequencies above 2 GHz in modern 7nm processes.
Conclusion
relOBI represents a significant advancement in radiation-hardened interconnect design. By strategically combining TMR and ECC, it provides full protection against single-bit faults with significantly less area overhead than traditional triplication methods. This makes it an ideal candidate for high-performance, low-latency microcontrollers operating in high-radiation environments. Future work includes physical radiation testing on a complete SoC and further optimizing ECC schemes for non-critical data.
Summary updated