Skip to content

Inside a 1 Gbit/s hardware entropy pipeline

A hardware entropy pipeline has four stages: a physical source, digitisation, health testing, and conditioning that compresses raw output into full-entropy bits. Throughput at the output is set by the entropy per raw sample, not by the sampling rate, which is why the compression ratio is the number that governs the design.

The stages, and what each is for

The source produces an analogue signal whose unpredictability follows from physics. Digitisation samples it. Health tests check continuously that the source is still behaving the way the entropy model assumes. Conditioning compresses the digitised stream so that the output carries close to one bit of entropy per bit.

Each stage has a failure mode the next one cannot fix. A source operating outside its model produces samples that look fine and carry less entropy than assumed. Digitisation that clips or saturates discards unpredictability silently. Conditioning applied at the wrong ratio produces output that passes every statistical test while carrying less entropy than claimed, because a good compression function makes under-entropic input look excellent.

Why the compression ratio governs everything

Raw samples never carry full entropy. A digitised phase-diffusion signal might carry a fraction of a bit of min-entropy per bit of ADC output, with the rest being classical noise and deterministic structure. Conditioning has to compress by at least the inverse of that fraction, with margin.

So output rate is the sampling rate divided by the compression ratio, and the compression ratio comes from the entropy estimate, and the entropy estimate is conditional on the model. A vendor quoting an output rate without saying what min-entropy assumption it rests on has quoted a number that cannot be checked. The measure and its conditionality is the pillar-2 treatment of this.

Health testing is the part that is usually thin

SP 800-90B requires continuous health tests, and the specified ones — repetition count and adaptive proportion — catch catastrophic failure. They do not catch a source that has drifted to two-thirds of its modelled entropy, because that source still produces output with no repeated values and no obvious bias.

Detecting that requires monitoring the physical quantities the entropy model depends on, not just the statistics of the output. That is the difference between a source that is tested and a source that is observed, and it is the whole of the Level 3 to Level 4 distinction.