Fault Tolerance¶

A classic argument for stochastic representation: information is spread across \(N\) bits, so many soft errors become small density errors instead of catastrophic MSB failures.

Soft-error intuition¶

Flipping \(k\) bits in a unipolar stream changes the decoded value by at most \(k/N\). For large \(N\), random SEUs look like extra variance — often cheaper to ride through than to harden every binary register bit.

graph LR
    SEU[Soft error flips] --> Bin[Binary datapath: MSB risk]
    SEU --> SC[SC datapath: ~k/N density shift]
    Bin --> Crash[Large value jump]
    SC --> Small[Graceful degradation]

What reliability papers typically compare¶

Axis	Questions to ask
Fault model	Stuck-at, SEU rate, aging
Metric	Output MSE vs application BER / accuracy
Cost	Extra SNG / redundancy energy
Baseline	Binary vs SC at iso-quality

SC is not automatically “more reliable” — poorly correlated or too-short streams can fail task metrics even with soft errors ignored. Reliability must be measured end-to-end.

Overlap with other taxonomies¶

Parallel SC multipliers for BNNs often discuss error resilience alongside energy.
Beyond-von-Neumann bitstream processing surveys position SC inside broader unreliable-device contexts.

📚 Primary literature (DOI only)¶

Circuit reliability: SC vs binary computing: DOI 10.1109/TCSII.2020.2993273
SC in beyond von-Neumann era (bitstream processing): DOI 10.1109/TCSII.2022.3161995
Energy-efficient binary-stochastic multipliers: DOI 10.1109/TCSII.2022.3210035
Cost/speed co-optimized parallel SC multiplier (BNN context): DOI 10.1109/TCSII.2025.3562199

Domain browse: Paper Map → Fault Tolerance & Reliability.