A memristor pulsed on the microsecond scale answers unrepeatably. LoMaRe's circuit turns that answer into a tunable distribution generator — the sampler that probabilistic compute has been imitating in software.
Keys, nonces and signatures are only ever as strong as the entropy behind them. brAIn is a drop-in hardware entropy source — the block that sits at the root of every secure chip. Certifiable against the standards buyers already demand (NIST SP 800-90B, AIS-31), in a standard chip process at room temperature, it integrates into secure elements, MCUs and HSMs — at orders of magnitude lower energy per bit than existing entropy circuits.
Confidence costs compute, so most edge devices skip it — and act on guesses they cannot grade. Modern policies generate each action by sampling, step by step; that sampling is exactly what brAIn does natively — accelerated, not emulated. Knowing when it does not know becomes affordable: many more possibilities per watt, or the same accuracy on a fraction of the power.
A robot in a field or a human space cannot be pre-programmed. Working out where it is, it tracks thousands of possible positions at once from noisy sensors, every cycle. Deciding what to do next, it tests thousands of candidate moves per second to pick one that will not hit anything. With brAIn every one of those possibilities is a device, not an arithmetic workload — more options explored in the same budget, better choices in clutter.
Edge to datacenter: a compute-in-memory architecture with randomness and compute in a single array runs sampling-bound AI natively — Monte-Carlo, Bayesian and generative sampling, not emulated — at a fraction of a GPU's energy. It targets the memory-bound decode stage of inference, where the industry runs memory-light accelerators, and is delivered as licensable IP or silicon (PCIe accelerator cards).
The feedback parameters set the mean and the variance of the output. Three parameter sets, three distributions, one device — and no arithmetic anywhere in the loop. This is the distribution the analogue belief accumulates from, produced by physics.
Samples never leave the analogue domain — the posterior accumulates in analogue, not in registers. A comparator at femtojoule class crosses the boundary once per decision — 256 to 1,000 samples — not once per sample. The digital domain is unchanged: host compute reads the belief, sets a policy, takes an interrupt; the imaging chain wakes only when the belief crosses.
brAIn keeps a live probabilistic belief about the world in the analogue domain, ahead of the converter — and hands the digital side one interrupt per decision, not one per sample.
Binned analogue signal, ahead of the converter.
Sampler array · conditioning · calibration · analogue posterior accumulator.
A comparator, at femtojoule class.
MCU · CPU · GPU · NPU · DSP. Reads belief, sets a policy, takes an interrupt.
Wakes only when the belief crosses the boundary.
One IP platform, three applications — the same primitive that secures a key can weigh a robot's next move and run sampling-bound inference without the energy bill.
A drop-in hardware entropy source — the block at the root of every secure chip. Certifiable against NIST SP 800-90B and AIS-31; standard process, room temperature; integrates into secure elements, MCUs and HSMs at a fraction of the energy per bit.
Talk to us about TRNG →Perception, planning and policy sampling for robots, autonomy stacks and humanoids — machines that must act under uncertainty, on a power budget. Many more possibilities per watt; knowing when it does not know becomes affordable.
Pilot programmes →A compute-in-memory accelerator that runs sampling-bound AI natively — Monte-Carlo, Bayesian and generative sampling, not emulated — for quant desks, national labs, pharma and cloud AI. Licensable IP or silicon.
Licensing →Probabilistic workloads are expensive on conventional hardware because randomness has to be manufactured — a PRNG, floating-point arithmetic and memory traffic for every sample. A memristor pulsed on the microsecond scale answers unrepeatably; brAIn calibrates that physical randomness into a controllable sampling operation, so the posterior can be accumulated by physics instead of computed by logic.
Because the belief lives in analogue, always-on sensing costs what a pulse costs — microwatts, continuous — and the expensive digital stages stay asleep until a decision is actually due. The digital host is untouched: it receives a belief and an interrupt, and runs whatever model it already runs.
Talk to us about brAIn — evaluation, licensing, or a joint development programme.
Get in touch