brAIn · analogue belief-gated sensing · Imperial College London spin-off

Introducing brAIn — the missing piece for edge AI.

Real Randomness, generated in hardware!

LoMaRe builds solid-state silicon that draws probability — and memory — directly from device physics. The belief is maintained in analogue; the digital domain is unchanged. Alongside: ultra-low-power PMRAM memory and SiMO near-infrared photodetection, all CMOS-compatible.

SCROLL · LIFT THE LID · SENSOR TAP → brAIn → DECISION
Heat spreaderlid
Host computedigital · unchanged · reads belief · interrupt
Conditioning · calibrationanalogue posterior accumulator
1T1R sampler arrayM memristive generators · sensor tap
Decisiona comparator · femtojoule class
watching · µW
pulses 0
bits 0
P(1) 0.50
ȳ 0.000 · σ 0.000
scroll to open
How it works · live

The belief is maintained in analogue. The digital domain is unchanged.

Samples never leave the analogue domain — the posterior accumulates in analogue, not in registers. The boundary is crossed once per decision — 256 to 1,000 samples — not once per sample. That is the whole architecture. Everything below is computed by a circuit simulation this page is running right now.

analogue posterior · liveWATCHING · µW
DECISION BOUNDARY POSTERIOR P(EVENT | OBSERVATIONS) t →
belief
0.12
samples this decision
0
decisions · interrupts
0
01 · Sensor tapanalogue · ahead of the converter

Binned analogue signal, before the ADC

The sensor is tapped in the analogue domain and the sampler array is pulsed against it. A stress pulse deforms the memristor; a sense read measures what it became — unrepeatable on the microsecond scale. Physics is the sampler.

sampler array · stress / sense · fig. 2B
02 · brAInsampler · conditioning · calibration · accumulator

The posterior accumulates in analogue — not in registers

Conditioning subtracts the previous read so drift cancels and the randomness survives; calibration holds the distribution where it is set; the analogue posterior accumulator integrates the evidence. No sample is ever digitised.

y(n) = x(n) − x(n−1) · threshold t(n) · bit stream
03 · Decisiona comparator · femtojoule class

One interrupt per decision, not per sample

A comparator crosses the boundary once per decision — every 256 to 1,000 samples. Host compute (MCU · CPU · GPU · NPU · DSP) is unchanged: it reads the belief, sets a policy, takes an interrupt. The imaging chain or inference engine wakes only when the belief crosses.

belief
0.12gated
last decision at — samples · digital domain: idle
Our technologies

Three IP lines, one materials-science engine.

Two distinct device physics: a stochastic memristive sampler that keeps a belief in analogue, and a deterministic multiferroic memory. Both slot into existing CMOS manufacturing and win on energy and cost. Each technology carries its own applications.

From the lab

Measured on physical devices — these ones.

The numbers

Measured on the device, not marketed.

PMRAM figures from our own characterisation, caveats intact.

<0pJ/bit
Energy per bit

Electric-field switching, transient current only — no steady current.

0ns
Read / write · measured

Sub-nanosecond achievable in simulation.

1013cycles
Endurance

Estimated; 10⁸–10⁹ measured to date, roadmap to 10¹⁵.

0°C
Operating temperature

Industrial and automotive high-temperature stability.

0nm
Minimum device size

Finite-element analysis; no anisotropy-driven heating.

0+patents
Granted worldwide

PMRAM alone; 10+ more on SiMO photodetection.

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

Probabilistic workloads are expensive on conventional hardware for one reason: randomness has to be manufactured — a PRNG, floating-point arithmetic and memory traffic spent producing each sample, imitating something physics does for free.

Memory devices are already stochastic. LoMaRe's patented primitive calibrates that physical randomness into a controllable sampling operation — so probability becomes a native quantity in silicon. The same materials science yields our memory, our photodetectors and our sensors.

Timeline

Milestones at LoMaRe.

2017

Spin-off incorporated

Founded as a spin-off from Imperial College London (Company No. 10723076), in partnership with Imperial Innovations.

2022

Innovate UK Smart Grant — memory

Awarded for PMRAM memory chip design. Completed.

2023

Innovate UK Smart Grant — photodetectors

Awarded for flexible near-infrared photodetectors. On-going.

Now

SiMO manufacturability

Study with the Centre for Process Innovation (CPI), Durham, toward TRL6 and large-area yield.

Uncertainty is the next frontier of compute.

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

Get in touch