Memristive distribution generator · WO 2024/200771 A1

Randomness isn't computed here. It's measured.

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.

01 · The module

Three stages. One signal path.

A pulse source stresses an array of memristors. An analogue conditioning stage differentiates and thresholds the response. A digital Monte-Carlo engine weighs the samples. Nothing in the path computes a random number — it reads one.

LIFTING THE LID · 1T1R ARRAY · CONDITIONING · MONTE-CARLO ENGINE
02 · The array

M generators in a crossbar.

Each crosspoint is one transistor and one memristor — a 1T1R cell. Rows carry the pulse; columns read the current. M hardware generators sample in parallel, one per parameter in the model, and every one of them is a different device with a different distribution.

FIG. 6 · 1T1R × M · MEMRISTIVE DISTRIBUTION GENERATORS
03 · One cell

Stress it. Sense it. Compare two reads.

A large pulse deforms the device's internal structure; a small pulse reads what it became. Between 1 and 100 µs that read is irreproducible. The circuit never uses the value itself — it uses the difference between two consecutive reads, which cancels the slow drift and amplifies the randomness.

memristor current · stress / sense pulses · fig. 2B
Vpulse → SW₁ → RM → R₁ · FIG. 1A
04 · Conditioning

Drift out. Bits out.

A sample-and-hold keeps the previous read. Subtracting it is a high-pass filter: the wander of the device disappears, the noise survives. A comparator turns the difference into a bit, and a low-pass of the bit stream feeds back as the threshold — a closed loop that holds the mean where you set it.

y(n) = x(n) − x(n−1) · threshold t(n) · bit stream
I/V · sense current → voltage
A(z) · Σ⁻ · B(z) · delay, subtract, shape
C(z) · D(z) · feedback sets the threshold t(n)
05 · The distribution

A distribution you can dial.

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 proposal distribution a Monte-Carlo method needs, produced by physics.

set Aset Bset Clive: —
FIG. 8 · OUTPUT HISTOGRAM · 250 000 PULSES PER PARAMETER SET
06 · Inference

The physics samples. The engine weighs.

Prior samples come off the array. A digital stage weights each by the likelihood of the observed data, and the posterior expectation is a weighted sum. Importance sampling in hardware — the expensive part done by a device, the cheap part by logic.

E[f(β)] ≈ Σ wᵢ · f(βᵢ)wᵢ ∝ p(data | βᵢ) · one MDG per parameter · M in the array
Heat spreaderlid
Monte-Carlo enginedigital · weights · posterior
Distribution conditioninganalogue · A(z) B(z) C(z) D(z)
1T1R sampler arrayM memristive generators
Substratepulse source · I/O
Vpulse · SW₁stress / sense
RMmemristor
R₁sense → I/V
I/V
A(z)
Σ⁻x(n) − x(n−1)
B(z)
t(n) ⊳ bit
C(z) · D(z)feedback
p(β)output histogram
pulses 0
bits 0
P(1) 0.50
ȳ 0.000 · σ 0.000
STRESS · SENSE · COMPARE
scroll to open
Where it matters

Anywhere a system has to sample — and can't afford to simulate it.

A hardware distribution generator fits wherever randomness is on the critical path: sampling-heavy inference, dither, keys, and the always-on sensing that today's silicon can only duty-cycle.

Bayesian inference at the edge

"Our model is Bayesian. Our hardware is not."

Proposal samples from the array, importance weights in logic — posterior updates without the PRNG, the floating-point, or the memory traffic.

Dither for data converters

"We spend digital gates to make noise for the ADC."

A drift-free analogue random signal from one memristor and a comparator, stable enough to bias and tune.

On-the-fly key generation

"Seeding the PRNG is the weakest step in the chain."

True random bits from device physics, differentiated to remove drift, tunable to a target mean.

Stochastic computing blocks

"Bit-stream arithmetic needs bit streams we can trust."

A hardware-efficient primitive for stochastic computing: one cell, one comparator, one stream.

Always-on sensing

"Keeping the front-end awake is our entire power budget."

Continuous belief instead of duty-cycled classification — sampling that costs what a pulse costs.

Image processing & ML dither

"Random arrays for dithering are a memory problem before they're a maths problem."

Array dither generated in place, per device, per pulse.

Origin
Imperial College London, EEE
Intellectual property
PCT WO 2024/200771 A1
Malik · Papavassiliou
Peer review
IEEE Trans. Circuits & Systems I
Validation
Measured on physical devices · 250k-pulse datasets

Uncertainty is the next frontier of compute.

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

info@lomaretech.com