Memristor Board PCI

ANCHOR

A COTS Memristor Hardware Platform for Temporal Signal Classification

An open-source mixed-signal platform that classifies temporal sensor signals using a physical 16×4 array of commercial off-the-shelf memristors on a standard PCB. Weights are stored as programmable resistance states and the computation happens in the signal path. No custom fabrication, no foundry access, and no digital multiply–accumulate.

New to this? A memristor is a resistor whose resistance you can set electrically, and which keeps that value when the power is removed. That makes it a memory cell and a multiplier at once: apply a voltage and Ohm's law does the arithmetic where the number is stored, instead of shuttling it to a processor first. That idea is called in-memory computing, and hardware built to compute the way nervous systems do is called neuromorphic.

What it does

The array is built from commercial Knowm self-directed-channel memristors, sold as discrete parts with no foundry access required.

An encoder maps signal amplitude to one of 16 array rows. Each selected row is read through a voltage divider, producing four resistance-weighted class voltages; four RC leaky integrators accumulate them over time, and the class is whichever integrator ends highest.

Note on the architecture. The encoder selects at most one row per sample, so ANCHOR does not perform a simultaneous full-vector crossbar operation. Temporal memory comes from the integrator decay, not from parallel matrix–vector multiplication. This is a deliberate design choice: it is what lets a 16×4 array of individually addressable COTS devices classify time-varying signals.

Encoding is event-driven: a row activates when the signal enters its amplitude band and changes only when the signal crosses a threshold. A dead band around zero means an idle signal starts no computation. There is no clock.

Results

Both tasks run on the same 64 devices, reprogrammed between them. Gesture errors occur only within the BT/TB and LR/RL pairs; none crosses between vertical and horizontal gestures.

What these accuracies measure. Held-out windows, not held-out recordings. The gesture segments were cut from nine continuous recordings and every recording contributes windows to both training and test, so 87% describes separability on this hardware, not generalisation to a new session, performer, or lighting. Each ECG class is drawn from a single MIT-BIH record, so 85% is likewise not patient-generalisation accuracy.

Power figures are peak-current-based estimates of the memristor read paths only; they exclude the ADS, computer, oscilloscope and integrator circuits.

Where these numbers come from. They are the hardware measurements reported in the ANCHOR paper, taken under its protocol: 40 training signals per class, with 25 gesture signals and 15 heartbeats per class held out for the hardware test. This repository ships a larger curated dataset (436 gesture segments, 637 beats) with its own train/validation/test manifests, so re-running the notebooks will not reproduce these exact figures. The test counts match (100 and 60); the training pools do not.

Repository layout

ANCHOR/
├── Docs/
│   └── User_Manual.md              build it, program it, run it - start here
├── Hardware/PCB_Design/
│   └── Memristor_Board_DIP/        KiCad schematic + layout, Gerbers, BOM, 3D models
├── Software/
│   ├── Memristor_Array_Weights/    Python notebooks - derive the resistance matrix
│   └── System_Operation/           C++ programs - program the array, run inference
├── Dataset/
│   ├── Hand_Gesture_Dataset/       436 recordings, 4 classes  (CC BY 4.0)
│   └── ECG_Heartbeat_Dataset/      637 MIT-BIH beats, 4 classes  (ODC-BY 1.0)
├── Webpage/                        project page, relative image paths
└── Images/

Getting started

→ Read Docs/User_Manual.md. It covers the hardware, the components, assembly, and three tutorials that take you from raw sensor data to a class voltage on the oscilloscope.

The short version:

You will need

• Hardware: the ANCHOR PCB, four Knowm M+SDC Memristor 16 Discrete 32-DIP packages (what these devices are), an Analog Discovery Studio, a bench supply, an oscilloscope
• Software: Windows 10/11, Digilent WaveForms, Visual Studio with C++ desktop development, Python 3.11+ with numpy / pandas / matplotlib

Build the C++ side with Software/System_Operation/ANCHOR.sln (Release | x64).

Datasets

Both are included in full, with fixed train/validation/test splits that the notebooks assert are disjoint before doing any work. The test split is reserved for hardware measurement and is never seen during weight derivation. The splits are segment-disjoint rather than recording-disjoint; each dataset's NOTICE.md explains what that limits the results to.

ECG data derives from the MIT-BIH Arrhythmia Database; see electrocardiography for background on the signal itself.

Gesture classes, four movement directions across the sensing surface:

ECG classes, each drawn from a single annotated MIT-BIH record:

The gesture data was recorded by the ANCHOR team from a photovoltaic light-sensing surface. The ECG data is derived from the MIT-BIH Arrhythmia Database. See each dataset's NOTICE.md for attribution and required citations.

Licence: what you can actually do

Everything here is open, but the parts carry different licences because they are different kinds of work. In short: you can build this board, sell it, change it, and publish what you learn. The obligations are attribution, and sharing back changes to the board design itself.

This table is a plain-language summary for orientation, not legal advice. The licence files are what govern, and they win wherever this table is less precise.

Citing

A paper describing ANCHOR is under review. The citation will be added here once it is published. Until then, please cite this repository. Machine-readable metadata is in CITATION.cff.

The light-sensing surface used for the gesture task is described in:

H. Das, I. Fahad, S. N. B. Tushar, S. H. Alam, G. Buchanan, D. Scott, G. S. Rose and S. Swaminathan. "In-Sensor Motion Recognition with Memristive System and Light Sensing Surfaces." IEEE ISVLSI, 2024, pp. 192–197. doi:10.1109/ISVLSI61997.2024.00044

If you want the longer story

None of this is needed to build or run the platform. It is here for context, and because the history is genuinely interesting: a circuit element predicted on symmetry grounds in 1971, missing for thirty-seven years, and now sold by the tube.

Where the idea came from

• A readable history of the memristor, from TU Dresden. The best place to start
• Chua predicted the device from the symmetry of the four circuit variables, before anyone had built one. Chua 1971
• HP Labs identified a physical device matching the prediction, thirty-seven years later. Strukov et al. 2008

What people build with them

• The standard review of memristive crossbars for computing. Xia and Yang 2019
• Early demonstrations that a crossbar can classify. Alibart et al. 2013 · Prezioso et al. 2015
• Where the field has got to. Aguirre et al. 2024 · Schuman et al. 2022

The parts and data we used

• Knowm's self-directed-channel devices, the ones in this board. Overview · Datasheet
• The photovoltaic sensing surface behind the gesture task. Das et al. 2024
• The ECG source, free to download and well documented. MIT-BIH on PhysioNet · Moody and Mark 2001

Authors

Imran Fahad*, Danny Scott*, SNB Tushar, Garrett S. Rose, Sai Swaminathan^†

Department of Electrical Engineering & Computer Science, University of Tennessee, Knoxville, United States.
* Equal contribution. ^† Corresponding author, sai@utk.edu

Developed in the EPIC Lab at the University of Tennessee, Knoxville.

Contact

Questions about building or using the platform are welcome:

• Imran Fahad, ifahad@vols.utk.edu
• Sai Swaminathan, sai@utk.edu

For bugs or design-file issues, please open an issue on the repository.