Source-linked AI summary
4K-Memristor Analog-Grade Passive Crossbar Circuit
Hyungjin Kim, Hussein Nili, Mahmood Mahmoodi, Dmitri Strukov
TL;DR
The paper addresses spread in integrated metal-oxide memristor current-voltage characteristics, identified as a major problem for practical neuromorphic computing. It develops uniform 64×64 passive crossbar circuits and demonstrates conductance tuning and MNIST classification.
Problem
Spread in integrated metal-oxide memristor current-voltage characteristics is identified as one of the most important problems prohibiting practical neuromorphic-computing use.
Method
The paper develops uniform 64×64 passive crossbar circuits and demonstrates conductance tuning and ex-situ-trained classification using a single perceptron network.
Results
Conductance tuning achieved < 4% average error when programming 4K gray-scale patterns, while MNIST classification fidelity was within 2% of the software-modeled limit.
Takeaways & Limitations
The demonstrations support passive crossbar memories as a platform for analog conductance tuning and neuromorphic classification.
Takeaways & Limitations
Tuning error remained at unswitchable memristors, including devices stuck in the high-resistance state, and future work should address back-end development.
Abstract
from arXiv · showhide
The superior density of passive analog-grade memristive crossbars may enable storing large synaptic weight matrices directly on specialized neuromorphic chips, thus avoiding costly off-chip communication. To ensure efficient use of such crossbars in neuromorphic computing circuits, variations of current-voltage characteristics of crosspoint devices must be substantially lower than those of memory cells with select transistors. Apparently, this requirement explains why there were so few demonstrations of neuromorphic system prototypes using passive crossbars. Here we report a 64x64 passive metal-oxide memristor crossbar circuit with ~99% device yield, based on a foundry-compatible fabrication process featuring etch-down patterning and low-temperature budget, conducive to vertical integration. The achieved ~26% variations of switching voltages of our devices were sufficient for programming 4K-pixel gray-scale patterns with an average tuning error smaller than 4%. The analog properties were further verified by experimentally demonstrating MNIST pattern classification with a fidelity close to the software-modeled limit for a network of this size, with an ~1% average error of import of ex-situ-calculated synaptic weights. We believe that our work is a significant improvement over the state-of-the-art passive crossbar memories in both complexity and analog properties.
Device Fabrication
The paper develops a larger passive crossbar using a modified, foundry-compatible fabrication process designed to improve uniformity, yield, and scalability. Etch-down patterning, CMP, conductive electrodes, and low-temperature processing support the resulting 64×64 array.
- Materials and uniformity: Thermal annealing adjusted the oxygen-vacancy profile, lowering virgin conductance and forming voltages while improving integration uniformity.The Al2O3 layer additionally suppressed small-voltage currents and line-to-line leakage, improving adjustable-current dynamic range.
- Materials and uniformity: Aluminum electrodes, TiN capping, and optimized planarization were used to improve line conductance, device uniformity, and yield.Aluminum provided better conductivity, while TiN prevented aluminum oxidation and CMP improved top-electrode quality.
- Process and scalability: Reactive-ion etching and CMP improved electrode geometry, topology, and isolation by reducing sidewall residue and lift-off artifacts.The fabricated array showed smooth top-electrode topology and no noticeable sidewall residue between electrodes.
Device Characterization
Device measurements characterize switching thresholds, current behavior, and conductance tuning across the 64×64 passive array. Despite threshold variation, most switchable devices were tuned accurately, with low average error.
- Current behavior: 2 µA-to-50 µA dynamic current range was measured at 0.25 V, with 200 µA-to-400 µA reset and set switching currents.Small-voltage I(V) behavior was balanced, and average nonlinearities were approximately 2 at 1 V and 1.1–1.3 at 0.25 V.
- Switching thresholds: 1.19 V and –1.39 V are the average set and reset threshold voltages, with standard deviations of 0.31 V and 0.37 V, respectively.The normalized threshold variation was approximately 26%.
- Conductance tuning: ~98% of switchable devices were tuned within 5% error, while the average absolute tuning error was ~3.76%.The reported average error was 2× better than in Ref. 21.
- Conductance tuning: The tuning experiment stopped at the desired 5% error, although 1% error was demonstrated for a device and a portion of the crossbar.Whole-crossbar tuning used multiple cycles to retune devices disturbed during earlier cycles.
MNIST Benchmark Classification
The passive crossbar was evaluated as the vector-by-matrix multiplication core of an ex-situ-trained, single-layer MNIST perceptron. Measured classification closely followed software predictions, while errors arose from unswitchable devices and narrow class margins.
- Classifier operation: The 64×10 crossbar performed vector-by-matrix multiplication, with the output row carrying the largest current determining the digit class.Neuron functionality, including bias, was emulated in software.
- Classification results: 1.87% behind the ideal software model was the experimental fidelity for the most accurate weight import.Measured fidelity was compared with software performance across 1%-to-50% weight-import errors.
- Classification results: Pattern “7” was correctly classified, whereas pattern “9” was misclassified partly because of large tuning errors at unswitchable memristors.The “9” example also had narrow current margins between the correct class and the closest competing classes “0” and “8”.
Summary
The work addresses current-voltage variability in passive metal-oxide memristors by demonstrating a uniform 64×64 crossbar with foundry-compatible fabrication and analog computing results. It achieves low-error conductance programming and MNIST fidelity close to software modeling, while future work targets integration and lower operating currents.
- Motivation: Current-voltage variation in integrated metal-oxide memristors is identified as a major problem for practical neuromorphic-computing use.The paper frames reducing this spread as its general focus.
- Contributions: Almost 99% working crosspoint memristors were integrated in a uniform 64×64 passive crossbar using a foundry-compatible process suitable for back-end-of-line or 3D integration.The paper presents this as a specific contribution.
- Analog demonstrations: < 4% average error was achieved when programming a 4K gray-scale pattern.This was experimentally demonstrated as a conductance-tuning result.
- Analog demonstrations: Within 2% of the software-based model was the classification fidelity of an ex-situ-trained, down-sampled 8×8 MNIST benchmark implemented in a single perceptron.The demonstrated network used the passive crossbar for the classification operation.
- Significance and outlook: 2× smaller tuning error in a crossbar with 10× more integrated devices than prior published work is presented as a significant step for passive crossbar memories.The comparison is stated as the authors’ assessment of the work.
- Significance and outlook: Future work should address back-end-of-line integration with conventional semiconductor circuits and reduce memristor operating and write current.These targets are stated as necessary for the technology’s ultimate performance.
Methods
The fabricated crossbar was wire-bonded, mounted on a custom printer board, and electrically characterized using computer-controlled Keysight instrumentation and custom C code.
- The crossbar array was wire-bonded and inserted into a custom printer board for testing and application demonstration.
- Keysight B1500A measurements and a 34980A switching matrix steered connections to crossbar inputs and outputs.
- A personal computer controlled the parameter analyzer and switching matrix through general-purpose interface and USB connections using custom C code.
one particular device are highlighted.
The study characterizes switching, tunes conductances across the crossbar, evaluates MNIST classification, and documents fabrication and switching-voltage correlations.
- Switching statistics: Switching thresholds were defined as the smallest absolute voltage causing a 20% conductance change measured at 0.25 V.
- Conductance tuning: Conductance tuning used automated write-verify pulse sequences, with tuning error defined from target and actual currents at 0.25 V.The reported tuning-error expression is 100×[Itarget(0.25V)-Iactual(0.25V)]/Itarget(0.25V).
- Conductance tuning: Devices were tuned to conductance levels from 3 µS to 45 µS with 1% error.
- Fabrication and device behavior: Etch-based patterning, chemical-mechanical polishing, and etch-back formed a fabrication process described as suitable for back-end-of-line or 3D integration.Sidewall residue without planarization shortened bottom electrodes, while switching-voltage data showed significant set-reset correlations.
- MNIST classification: The MNIST evaluation used a 64×10 single-layer perceptron with rectify-linear neurons on 8×8 down-sampled images.The dataset comprised 60,000 training and 10,000 test images.
- Software-based training: Software training used backpropagation with a 0.01 learning rate, batch size 100, 50% dropout, and weights clipped to [0, 1].