Skip to content

The TRNG inside your SoC, what your silicon vendor does not tell you

Most systems already contain a TRNG — inside the SoC, provided by the silicon vendor, usually based on ring-oscillator jitter. It is rarely documented in enough detail to evaluate: the entropy model, the min-entropy estimate and what health testing runs are often not in any datasheet you can obtain.

The competitive set is not who you think

A discussion about entropy sources usually becomes a discussion about QRNG vendors. That is the wrong frame. The installed base is the TRNG already inside the SoC in the device you are shipping, and it is the source almost every design actually uses.

It costs nothing extra, it is already integrated, and it is on the die. Any argument for a different source has to be an argument against that one, on its merits.

How they usually work

Most embedded TRNGs derive entropy from jitter between free-running ring oscillators. Two oscillators drift relative to each other because of thermal noise in the transistors, and sampling one with the other captures some of that drift.

It is a reasonable mechanism. It is cheap in area, needs no special process, and the entropy is genuinely physical.

What is hard to establish

  • The entropy model. Jitter accumulation depends on oscillator design, supply noise, temperature and layout. Modelling it well is difficult, and the model is often not published.
  • The min-entropy estimate. Frequently quoted as a single number with no statement of what it is conditional on, and sometimes quoted post-conditioning, which makes it uninformative.
  • Health testing. Whether anything runs continuously, and what it would detect, is often undocumented.
  • Behaviour across the envelope. Ring oscillators are sensitive to supply and temperature, and the entropy can fall at the corners without the output looking different.

Why this matters more than it seems

Two oscillators on the same die share a supply and a substrate. Under some conditions they can pull toward each other — injection locking — which reduces relative jitter and therefore entropy, while producing output that still passes statistical tests.

It is a documented phenomenon in the literature and a plausible target for a deliberate attack on a physically accessible device. It is also exactly the class of failure statistical tests cannot see.

The question to put to your silicon vendor

Not "does it have a TRNG" but: what is the entropy model, what is min-entropy per raw sample and conditional on what, what health tests run, and what happens at the corners of the operating envelope? The full checklist applies, and the answers determine whether an additional source is worth its cost.