Natural Noise
Encryption with 10 times more entropy that fits an ultra-low power budget
Dear SoTA,
Before Living Proof started, we spent a long time thinking about randomness. If you want to implement cryptography on a new system, one of the foundational questions you need to answer is: where will your randomness come from? Secure communication protocols, certificate signing, and key generation all require a source of random bits to ensure instances are unique.
It takes a surprisingly large circuit to generate random bits in conventional CMOS, and even with thousands of transistors the quality is often unreliable. We use measures of information entropy to quantify how unpredictable a sequence of random bits is. Higher entropy means more unpredictable, and therefore more secure. A typical hardware random number generator yields roughly 0.1 bits of usable entropy per raw bit it emits. Producing 100 bits of cryptographic-grade randomness thus requires at least a thousand raw bits, which are then post-processed through algorithms whose computational cost scales with the input size. Stronger security means more computation, more power draw, and a longer wait.
On a laptop you never notice this. With plenty of space for dedicated hardware and optimized instructions, the resources required for cryptography are dwarfed by the system’s overall capacity. On a resource-constrained sensor asked to last ten years on a single battery, any additional overhead is a tough pill to swallow–even for something as important as security.
We started to wonder: how could you implement security primitives like random bit generation with minimal added resources? We found a possible solution in an existing form of nonvolatile memory. Inside every magnetoresistive RAM (MRAM) chip sits an array of magnetic tunnel junctions (MTJs). The fundamental physics of these devices comes from two magnetic layers separated by an insulator, each a nanometre thick. Send a current pulse through the stack, and the magnetization of the “free layer” flips to the opposite state with some probability, a genuinely stochastic event arising from the behaviour of quantum spins. The useful part is how clean that randomness is. Plot the probability of a flip against the duration of the pulse and you get a smooth, predictable curve rather than a scatter of noise, which means you can dial the odds of a bit-flip anywhere from near zero to near certainty. Instead of a messy, ephemeral random signal in a circuit, you get a tunable stochastic process whose final state is persistent. As a bonus, that state can be read with exactly the same machinery used for standard memory operations, and the “probabilistic write” requires only minor modifications of peripheral circuitry.
What we didn’t expect was just how much higher the entropy from this process would be. We found that MTJ-based TRNG could attain a min-entropy rate of 0.99 / bit, using a minimum of just one MTJ and less than one hundred transistors in an integrated design. Compared to CMOS ring oscillator designs which generate ~0.1 bits of min-entropy / bit using thousands of transistors, that translates to at least two orders of magnitude improvement in entropy rate per unit area at an equivalent process node. Even better, if a device already uses MRAM as part of its memory system, the area overhead to add TRNG is essentially zero.
The only sources that compete with the MTJ’s raw noise quality are dedicated quantum random number generators (QRNGs), which have seen increasing adoption as the post-quantum cryptography (PQC) transition increases entropy demands via new key derivation and signing algorithms. These high-entropy QRNGs require their own optical hardware, converters, processor, and I/O; adding up to massive form factors from PCIe card up to 1U server size. The MTJ gives you the quality of the quantum device with the energy budget of the embedded one, on-chip and at picojoules of power consumption per bit.
But validating the device physics doesn’t mean the solution is complete. Manufacturing millions of these cells introduces tiny variations from one MTJ to the next, and running them deliberately in their stochastic regime, flipping back and forth rather than holding a stored value, pushes on their endurance to a degree memory workloads rarely approach. Before this becomes a product, device-to-device variation and lifetime under stochastic operation have to be characterised properly. We treat that as engineering work to be done to ensure first principles can translate to real impact.
Consider where efficient, reliable entropy actually changes the workflow. A smart meter expected to run untouched for a decade, rotating keys autonomously for best-practice security without breaking its power budget. Or a medical implant that has to authenticate every reading it sends, where every microjoule spent on cryptography is a microjoule taken from the battery keeping someone alive. These are real constraints today, and the missing foundation for both is a practical entropy source.
There is a deeper idea here that we are still working out, and it came from biology rather than physics. A cell is an island of order in a sea of chaos, and it stays that way by exporting entropy rather than fighting it directly. Eat an apple and your body takes in an ordered, low-entropy structure and ejects disorder as heat and waste. Order is maintained by getting rid of disorder efficiently. Life also does something subtler that our friend Kimia Witte calls cross-level stochastic-to-deterministic collapse. At the membrane, millions of ion channels fire stochastically every microsecond. Those collapse into a handful of structural changes in the cytoskeleton each millisecond. Those collapse again into a single gene-expression decision over minutes. Countless chaotic events at one scale funnel into one definite outcome at the scale above. The retina takes in something like gigabytes of photon data a second and the conscious mind downstream of it runs at roughly ten bits a second, the speed of a fast typist. Almost all of that information goes somewhere, and classical information theory, built to carry a message intact from one point to another, does not really describe where.
This matters to us for a concrete reason. The same MTJ we use to generate maximum information entropy depends on holding its physical structure at minimum entropy, a tunnelling barrier kept free of the conductive defects that would corrupt the crystal. Generating disorder cleanly requires maintaining order somewhere else. Life has been solving exactly that trade-off for billions of years, across scales, under tight energy budgets, which makes it the existing proof that the entropy problem has good solutions and not just expensive ones. Kimia’s work on Computation as Organisation treats information not as content to be preserved but as a structure of relationships that guides how a system behaves, and aims at a mathematics of those relational invariants that holds for silicon and cells alike. We think that language is what the field is missing.
We are mapping insights like these into an Outcomes Graph, working from the source of randomness outward to the systems that have to live within an energy budget while staying secure. If you work on MRAM, true random number generation, edge cryptography, or natural computation, we would like to hear from you.
Help us build natural noise and “cell crypto”.
Thane Campbell & Shannon Egan
Living Proof
Thane Campbell is Dean at Deep Science Ventures leading partnerships for the world’s 1st PhD in invention — The Venture Science Doctorate. His PhD Fellows are launching companies with new Fusion reactors, virtual power plants, scalable elastocaloric cooling, clean batteries and epigenetics to give 2 billion women 5 more years of health.
Shannon Egan is the founder of a memory technology startup incubated at Deep Science Ventures, where she leads R&D and product. Her multidisciplinary research background spans neuroscience, particle physics, quantum materials, and magnetic devices for computing.