Analog compute in memory

Large AI models, on a robot's power budget.

P2Matrix chips do the multiplying where the AI model is stored. The model's weights never travel, so a robot can run a large language or vision model on a few watts, without the cloud.

1 dieholds the 4-bit weights of a 27-billion-parameter model (Candidate F, 28 nm, modelled)
< 3 Wper weight die with every array busy (modelled)
0 bytesof model weights moved per token: only small inputs and results travel
10U.S. provisional patent applications filed since October 2026

All P2Matrix performance figures on this site are pre-silicon estimates from our models. Nothing has been fabricated yet.

The problem

AI chips spend their energy moving the model, not thinking with it.

A language model is billions of numbers, its weights. To write each word, the chip multiplies the input by every weight in the model. In a GPU those weights sit in memory chips beside the processor and are fetched across again for every word.

Think of a chef who must walk to the warehouse for every ingredient, for every dish. The cooking is quick; the walking is what costs time and energy. Fetching a weight from memory takes far more energy than multiplying with it. Data centres pay for this walk with power plants and liquid cooling. A robot pays with its battery and its heat budget, or by sending its thinking to the cloud.

That is why today's robots either run small models or carry a 40 to 130 W GPU computer, and why so many depend on a network connection to think.

Today: the weights travelFor every token, the model is fetched from memory again.memory stacksprocessor
Our approach

Compute where the weights live.

P2Matrix: the weights stay putCells multiply and add the inputs where the weights sit.inputs indot products out

P2Matrix stores each weight as an analog value inside a memory cell, and lets the physics of the cells do the multiply-and-add, the dot product at the heart of every AI model.

Inputs arrive as electrical signals on one set of lines. Every cell they cross contributes in proportion to the weight it holds, and the contributions add up on the other set of lines. A whole matrix of weights is multiplied by a vector of inputs in one step, in place. The weights are written once when a model is loaded and then only read.

The memory is non-volatile: it keeps the model with the power off, so a robot wakes up ready and spends nothing to keep its model alive.

⇆

Weights never move

Only small input vectors and results travel between chips: kilobytes per word, not gigabytes.

⚡

Low power

No memory bus to drive and nothing to refresh. The array draws power only while it computes.

▦

Dense

Weight memory is built in layers, so one die holds a model that needs a GPU's whole memory today.

⚙

Standard manufacturing

Our lead design is aimed at existing foundry processes at 130 nm and 28 nm, not leading-edge nodes.

The honest part: analog accuracy

Analog computing is fast and frugal, but it is not exact. Cells vary, values drift and every reading carries noise. We design for that rather than hide it: weights use 4 bits, as many models now ship; each weight is written and then checked; the chip calibrates itself against known reference values; and the read is sized for the precision the model needs. Our models of noise, drift and variation are in the investor data room, with the open questions we still have to settle in silicon.

How we differ

Different from GPUs, wafer-scale chips and other in-memory designs.

Every AI chip has to answer one question: where do the weights live, and how far do they travel? Here is how the main approaches answer it, as described in each company's public material.

ApproachExamplesWhere the weights liveWhat it is built forThe trade-off
GPUNVIDIAHigh-bandwidth memory beside the processor; fetched for every tokenTraining and inference in data centres; Jetson modules for robotsFlexible and fast, but the memory traffic costs power: robot modules are rated 40 to 130 W
Wafer-scaleCerebrasOn-chip SRAM across a whole silicon waferVery fast data-centre inference and trainingA wafer-sized system with data-centre power and cooling
SRAM inference chipsGroq, Etched (dataflow / model-specific)Fast on-chip memory, spread over many chipsData-centre tokens per secondOn-chip memory is small, so a large model needs many chips or racks
Digital in-memoryd-Matrix, AxeleraSRAM with digital arithmetic inside or beside itData-centre inference; edge visionExact digital math; capacity per chip limited by SRAM
Analog in-memoryMythic, Sagence, EnCharge AI, TetraMemFlash cells, capacitors or resistive cells that compute in analogEdge vision and audio; data-centre inferenceVery efficient; analog precision has to be managed
P2MatrixCandidates F and JNon-volatile analog weight memory, built in layers, gigabytes per dieRobot-centric language, vision and action models at a few wattsPre-silicon; analog precision managed by write-verify and calibration

Company names and product names belong to their owners. Descriptions are our summary of public information as of November 2026 and may be incomplete.

Built for robots

A robot should think on board.

Robots are moving from scripted motions to models that see, talk, plan and act: vision-language-action models, perception, and language models that reason about the task. Those models are large, and a robot has a battery, a heat budget and no guarantee of a network.

  • Battery and heat: a weight engine that draws a fraction of a watt leaves the battery for the motors.
  • Latency: decisions are made on the robot, not after a round trip to a server.
  • Privacy and reliability: no camera feed has to leave the robot, and nothing stops when the Wi-Fi does.
  • Learning in the field: small adapters update the model on the robot without rewriting its weights, so it learns without forgetting.
  • Instant on: the model stays in the chip with the power off.
Built for robotsPower to run a 27B model at about 20 tokens/sGPU robot computer, published module rating40-130 WP2Matrix F, robot configuration, modelledabout 0.3 W0.1 W1 W10 W100 WLog scale. Ours: modelled weight engine and processor while answering.GPU: whole-module rating. Not a like-for-like benchmark.
Design candidates

Two lead designs, one architecture.

We explored many memory technologies under one architecture and one reference model, so they could be compared on the same terms. Two lead today.

Lead design130 nm and 28 nm

Candidate F: analog weight decks on standard CMOS

Thin layers of analog weight memory built in the wiring above ordinary logic. Made on foundry lines with open capacity, starting at 130 nm and moving to 28 nm for more memory per die. Rewritable, so a robot can take a new model after it ships.

  • About 26 GB of 4-bit weights per 28 nm die (modelled)
  • Fractions of a watt to about 2 W per die
  • First test silicon planned at 130 nm
Capacity designWith a memory partner

Candidate J: dense 3D memory as the multiplier

The cells of a production 3D memory hold the weights as analog values, with a bonded logic wafer reading the results. About four times the weights per die of Candidate F, for the largest models. Built with a memory maker.

  • About 113 GB of 4-bit weights per die (design study)
  • Trillion-parameter-class models in one package
  • Read speed is the open engineering question
How we measure

Numbers you can check.

Dot products per second

At chip level we count multiply-adds per second for a whole matrix times a vector, the unit every AI model is built from, so the comparison is not muddied by memory buses or caches.

Like against like

One user against one user, a fully loaded chip against a fully loaded GPU, and the same amount of memory on both sides. We never set a full-load figure beside a one-user figure.

Labelled by evidence

Every figure says whether it is modelled, simulated, published or measured. Today all of ours are modelled or simulated; measured numbers will replace them as silicon arrives.

Roadmap

From models to silicon.

2026

Architecture and IP

Design studies across memory technologies, circuit simulation, 10 U.S. provisional applications.

2027

Test silicon

130 nm test chips and cell test structures: the first measured dot products.

2028

Product die

28 nm weight die and a robot module development kit.

In parallel

Memory partner

Candidate J with a 3D memory maker, for the largest models.

Plans, not commitments; timing depends on funding and foundry schedules.

Company

The team.

P2Matrix is a startup building analog AI compute for robots. We work in the open with our investors: our data room is the company's lab notebook, failed paths included.

photo

Name

Founder and CEO

Short bio to come.

photo

Name

Co-founder, chief architect

Short bio to come.

photo

Name

Advisor, semiconductor manufacturing

Short bio to come.

photo

Name

Advisor, robotics

Short bio to come.

Our AI engineering team

Our engineers work with AI models from several vendors that draft, check and review each other's work, each under its own name and persona. They will be introduced here.

face

Name

AI member: design
face

Name

AI member: review
face

Name

AI member: simulation
face

Name

AI member: research

Careers

We will be hiring in analog and mixed-signal circuit design, device physics and ML compilers. Tell us what you would build.

contact@p2matrix.ai

Partners

Foundries, memory makers, robot builders and universities: we want to hear from you about test silicon, models and use cases.

contact@p2matrix.ai

News

Patent filings, test-chip milestones and papers will be posted here.

Nothing to announce yet.

Investor data room

The full design studies, simulations, patent applications, cost models and open problems, with an assistant that answers questions about every page. Access by invitation, under NDA.

Enter the data room →