$\mu $ RNG: A 300–950 mV, 323 Gbps/W All-Digital Full-Entropy True Random Number Generator in 14 nm FinFET CMOS

This paper describes μRNG, an ultra-lightweight all-digital full-entropy true-random number generator (TRNG), fabricated in 14 nm high-k/metal-gate FinFET CMOS, targeted for on-die generation of cryptographic keys in energy-constrained IoT and wearable platforms. The μRNG combines the entropy of multiple independent sources to generate an output bitstream that is indistinguishable from an ideal unbiased entropy source. Three independent self-calibrating all-digital entropy sources, coupled with XOR feedback shift-register based correlation suppressors and an in-line compact Barak-Impagliazzo- Wigderson (BIW) extractor enable ultra-low energy consumption of 3 pJ/full-entropy bit with a dense layout occupying 1008 μm², while achieving: (i) 162.5 Mbps full-entropy throughput at 1.3 GHz operation, with total power consumption of 1.5 mW and leakage power component of 90 μW, measured at 0.75 V, 25°C, (ii) mutually uncorrelated raw bitstreams from the three entropy sources with phi-coefficient cross-correlation <;0.003, (iii) extracted full-entropy bitstream that passes all 16 NIST RNG tests with measured Shannon entropy up to 0.9999999995, and lower-bound min-entropy H_∞ > 0.99, (iv) hysteresis-free extracted output for lags 1-1000, with ACF ~0 within 95% confidence bounds of a Gaussian distribution (μ = 0, σ² = 0.002), (v) wide operating supply voltage range of 300-950 mV with throughput scaling to 225 Mbps at 950 mV and robust subthreshold voltage performance of 400 Kbps, 4 μW, measured at 300 mV, 25°C, (vi) peak energy-efficiency of 323 Gbps/W at near-threshold voltage of 400 mV, with full-entropy throughput of 8.6 Mbps, total power consumption of 27 μW, (vii) 6.5× reduction in gate count and 5.4× lower energy consumption compared to conventional AES-based entropy extractors.

I. Introduction

Full-entropy true random number generators (TRNGs) are foundational primitives of secure platforms, providing independent and identically distributed keys for data encryption, secure wireless communications and media content protection [1]. A full-entropy cryptographic-quality bit key is obtained from a bitstream where every bit pattern has uniform occurrence probability approaching over a large infinite dataset, with min-entropy . (Min-entropy is computed as the negative logarithm of the probability of the most likely pattern). These random bitstreams are generated using entropy harvesters that sample and digitize entropy from physical sources such as thermal noise, random telegraph noise or device oxide breakdown. Among natural sources, the technology-invariance and frequency-independence of thermal noise makes it the most popular source of randomness in TRNG designs [2], [6], [9], [10], [19]–[21]. Conventional entropy sources employ high-gain differential amplifiers to amplify differential thermal noise voltage sampled across a resistor-pair, followed by digitization using an analog-to-digital converter [7], [9]. Alternate sources of randomness include Si nano-devices [11], biased inverters [24], oxide breakdown [12], [13] or random telegraph noise [8]. While these approaches are effective at producing high-entropy bitstreams, they also present significant challenges when ported to high volume manufacturing environments in advanced sub-14 nm processes, due to stable analog supply voltage requirements and sensitivity to temperature and aging-induced device drifts. A separate category of entropy sources harnesses the entropy contained in phase-noise by sampling edge-jitter of ring oscillators [3], [4], [14]–[16], coupled oscillators [5], [7], [31], [34] or Fibonacci/Galois ring oscillators [17], [18] using phase-detectors. Another popular approach to harvesting thermal noise employs the metastable behavior of cross-coupled inverters, latches and SRAM bitcells [19]–[23].