Number Generators (RNG / SNG)

In stochastic computing, entropy and encoding often dominate area and energy. A beautiful one-gate multiplier loses if every operand needs an expensive stochastic number generator.

graph TD
    Src[Entropy / sequence source] --> Cmp[Compare with binary x]
    Cmp --> Bit[Bitstream X_t]
    Src --> Types{Source class}
    Types --> LFSR[LFSR / PRNG]
    Types --> LD[Low-discrepancy: Sobol / Halton]
    Types --> TRNG[Device TRNG / MTJ]
    Types --> Mem[Memory-compatible B2S]

---

Roles

Block	Job
RNG	Provide (pseudo)uniform bits / words
SNG / B2S	Map binary value \(x\) → density-\(x\) stream
Shared vs private	Trade area against correlation risk

---

Design axes

1. Statistical quality — period, pairwise independence, spectral artifacts
2. Area / power — LFSR bit width, sharing factor, TRNG amplifiers
3. Determinism — reproducibility for test; low-discrepancy for short \(N\)
4. Technology coupling — DRAM-friendly converters, spintronic TRNGs, CIM-friendly pulse codes

---

Practical rules of thumb

• Budget SNG energy per bit in every comparison against binary baselines.
• Never share one LFSR across both inputs of an AND multiplier unless you have proven the bias is acceptable.
• For short-\(N\) DSP / NN, evaluate deterministic sequences before lengthening streams.

---

📚 Primary literature (DOI only)

• Low-cost high-accuracy random number source for SC: DOI 10.1109/TCAD.2025.3585800
• DRAM-compatible deterministic B2S conversion: DOI 10.1109/TCSI.2026.3701420
• Spintronic tunable TRNG design: DOI 10.1109/TCAD.2024.3446715
• Probabilistic CIM for MCMC: DOI 10.1109/TCSI.2023.3337529

Domain filter: Paper Map → Number Generators.