Probabilistic Compute · Born at the frontier of physics and inference

Continuous reasoning, at microwatts.

Lomare builds silicon that draws probability directly from device physics — maintaining a live belief about the world at a small fraction of the energy digital hardware spends imitating the same thing. From always-on sensors to inference at datacenter scale.

prior belief updated belief
The Status Quo

Always-on AI isn't always on. It's asleep, guessing when to wake.

Every battery-powered intelligent device faces the same constraint: it cannot afford to run a neural network continuously. So the industry works around it — duty cycling, threshold triggers, heuristic wake-up logic. These reduce average power, but they buy that saving by giving up awareness.

Failure mode 01

Duty cycling creates blind intervals

Between scheduled wake-ups the device is not merely idle — it is unaware. Events that fall in the gap are simply never seen, and no amount of downstream model accuracy recovers them.

Failure mode 02

Thresholds are brittle

Fixed triggers respond to instantaneous amplitude, not meaning. They drift out of calibration with the environment, and must be tuned conservatively — which means waking on noise.

Failure mode 03

False wake-ups spend the budget

Each unnecessary escalation costs milliwatt-scale inference. In deployments where events are genuinely rare, most of the energy budget is consumed confirming that nothing happened.

Our Approach
Sampling isn't the bottleneck. Simulating it is.

Probabilistic workloads are expensive on conventional hardware for a specific reason: randomness has to be manufactured. A pseudo-random number generator, floating-point arithmetic and repeated memory accesses are spent producing each effective sample — picojoules to nanojoules of energy to imitate something physics does for free.

Memory devices are already stochastic. Their switching behaviour is probabilistic by nature, and conventional design treats this as a defect to be suppressed. Lomare's patented primitive does the opposite: it calibrates that physical randomness into a controllable sampling operation, so probability becomes a native quantity in silicon rather than an arithmetic result.

How It Works

Belief-gated inference: attention that costs almost nothing.

Rather than waking on a threshold, the system holds a continuously updated probabilistic belief about a latent variable, and escalates only when the evidence justifies it. Heavy inference still happens — just far less often, and for far better reasons.

Stage 1 · Watch

Continuous belief

A probabilistic filter ingests raw sensor input and maintains a posterior over the state of the world, updated incrementally with every observation. Never asleep, never blind.

~ microwatts, continuous
Stage 2 · Decide

Confidence gating

Escalation is driven by belief drift and uncertainty, not signal amplitude. Environmental noise is rejected at the belief layer, before any classifier is invoked.

evidence, not thresholds
Stage 3 · Act

Full inference on demand

When the posterior crosses a decision boundary, a downstream neural network wakes, resolves the event, and returns to deep sleep. The MDG is agnostic to the model it gates.

~ milliwatts, millisecond bursts
The Numbers

Measured in joules per posterior update.

The relevant unit for always-on inference is not operations per second — it is the energy cost of one belief update. That is where physical sampling separates from arithmetic emulation.

Digital probabilistic inference
PRNG, floating-point arithmetic, memory access
picojoule–nanojouleper effective sample
Lomare · first-generation silicon
Sampling from engineered device stochasticity
picojoule-classper effective sample
Lomare · proprietary NVM substrate
Sampling native to the memory device itself
femtojoule-classtargeted, per sample
Conservative estimates imply a 10³–10⁶× reduction in energy per sample relative to digital implementations, and a 10²–10³× reduction in total system energy on first-generation silicon for deployment classes where the monitored state spends the overwhelming majority of its operating life outside the classifier's region of interest. Figures are design targets supported by published device-level measurements; first-generation silicon is in development.
Where It Matters

Anywhere the world changes slowly — and matters suddenly.

Belief-gated silicon fits deployments where events are rare, energy is finite, and missing the moment is expensive.

Industrial condition monitoring

"We replace sensor batteries more often than the bearings they're supposed to be watching."

Multi-year deployments on a single cell, with no blind intervals between wake-ups.

Rail & transit maintenance

"Every jolt on the track sets our sensors off. Almost none of it means anything."

Escalation on accumulated evidence rather than vibration amplitude, on energy-harvested power.

Always-listening audio

"Keeping the microphone path awake is essentially our entire power budget."

Continuous acoustic awareness at a fraction of always-on feature extraction, viable in hearables.

Environmental sensing

"Our field nodes are duty-cycled so hard they miss the events we deployed them to catch."

Persistent attention instead of blind intervals, over multi-year solar or disposable deployments.

Wearable health

"Continuous monitoring means continuous charging — so people stop wearing the device."

Physiological awareness without the wake-on-every-frame penalty of current biometric front-ends.

Perimeter & asset monitoring

"Wind and rain generate far more alerts than anything we actually care about."

Environmental noise rejected in the belief layer, before the classifier is ever invoked.

The Trajectory

The same physics that wakes a sensor will reshape the datacenter.

The workloads defining modern AI — generative models, Bayesian methods, reinforcement learning, reasoning under uncertainty — are sampling-heavy at their core. Deterministic accelerators manufacture that randomness with arithmetic, and a growing share of AI's energy bill is spent doing so.

Hardware that samples natively changes the economics. Belief-gated sensing at the edge is the first product of a primitive whose larger destination is datacenter-scale probabilistic inference — sampling as a first-class operation in the fabric of AI compute, not an emulated one.

Edge sensing Probabilistic silicon Datacenter AI
Origin
Imperial College London, EEE
Intellectual property
PCT WO 2024/200771 A1
Peer review
IEEE Trans. Circuits & Systems I
Validation
Demonstrated on physical devices

Uncertainty is the next frontier of compute.

Partner, invest, or build with us — we'd like to hear from you.

info@lomaretech.com