pith. sign in

arxiv: 1906.12045 · v1 · pith:HEPMRFXUnew · submitted 2019-06-27 · 💻 cs.ET · cs.NE· physics.app-ph

4K-Memristor Analog-Grade Passive Crossbar Circuit

Hyungjin Kim , Hussein Nili , Mahmood Mahmoodi , Dmitri Strukov This is my paper

Pith reviewed 2026-05-25 14:23 UTC · model grok-4.3

classification 💻 cs.ET cs.NEphysics.app-ph
keywords memristor crossbarpassive arrayanalog neuromorphicMNIST classificationdevice variationfoundry processsynaptic weights
0
0 comments X

The pith

A 64x64 passive memristor crossbar programs 4K analog weights with under 4% error and classifies MNIST near the software limit.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper shows how a foundry-compatible process can build a large passive crossbar of metal-oxide memristors with 99% yield. Even with 26% variation in the voltages needed to switch the devices, the array can still be tuned to hold gray-scale patterns with average error below 4%. The same hardware then takes weights calculated in software and uses them to classify MNIST handwritten digits, staying within about 1% of the ideal software accuracy. This matters because it removes the need for extra transistors at every junction, which would otherwise cut density and raise power use in neuromorphic chips. The low-temperature steps also open the door to stacking layers vertically.

Core claim

The central claim is that ~26% switching-voltage variation in a 64x64 passive metal-oxide memristor crossbar is still low enough to program 4K synaptic weights accurately, as proven by gray-scale pattern storage below 4% error and MNIST classification with only ~1% average deviation from the ex-situ software model.

What carries the argument

The 64x64 passive crossbar circuit, which stores analog weights at each crosspoint without select transistors.

If this is right

  • The low-temperature etch-down process allows vertical stacking of multiple crossbar layers.
  • Ex-situ weight import can reach near-ideal accuracy for networks of this size without on-chip training.
  • 99% device yield makes the approach viable for building larger neuromorphic prototypes.
  • Analog tuning remains precise enough for real pattern-classification workloads.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • On-chip weight updates may still be limited by the same voltage variations even if ex-situ import works well.
  • The same crossbar approach could be tested with other datasets or network sizes to see where the variation limit appears.
  • Hybrid systems that combine this passive array with digital control logic could reduce off-chip data movement further.

Load-bearing premise

That the measured 26% variation and high yield will continue to support accurate programming when the array is scaled up or when weights must be updated on the chip instead of imported from outside.

What would settle it

Running the MNIST classification test on a 128x128 or larger version of the same crossbar and checking whether the error stays within 1% of the software model.

Figures

Figures reproduced from arXiv: 1906.12045 by Dmitri Strukov, Hussein Nili, Hyungjin Kim, Mahmood Mahmoodi.

Figure 1
Figure 1. Figure 1: Memristive crossbar array. (a) Scanning electron microscope image of the fabricated 64×64 memristor crossbar array. Top, bottom left, and bottom right insets show, correspondingly, zoom-in on the portion of the crossbar, layers at the device cross-section with corresponding thicknesses in nanometers, and packaged chip. (b) Representative as-fabricated and (c) after forming I-Vs, measured with a quasi-stati… view at source ↗
Figure 2
Figure 2. Figure 2: Switching statistics for crossbar devices. (a) Measured evolution of conductance upon application of increasing amplitude voltage pulses. All parameters of the utilized pulse sequences are similar to those shown in inset of Fig. 3c, except for 50 mV incremental step. (b￾d) Extracted statistics of switching threshold voltages, defined as a smallest absolute voltage at which device conductance measured at 0.… view at source ↗
Figure 3
Figure 3. Figure 3: Conductance tuning in the crossbar circuit. (a) The desired device conductances in the range of 10 µS to 100 µS, which corresponds to the gray-scale quantized Einstein image and (b) their actual measured values after completing automated tuning with 5% error. (c) Corresponding tuning statistics. Inset shows details of the write-verify pulse sequence. Tuning error is defined as 100×[Itarget(0.25V)-Iactual(0… view at source ↗
Figure 4
Figure 4. Figure 4: Pattern classification of MNIST images. (a) Portion of the crossbar circuit utilized in the implementation of 64×10 single-layer perceptron. (b) Examples of target and (c) actual conductances after tuning with 1% error. (d) Measured classification fidelity and its comparison with simulation results as a function of weight import accuracy. In each simulation trial, the weights were selected randomly from ra… view at source ↗
read the original abstract

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.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 2 minor

Summary. The manuscript reports fabrication of a 64x64 passive metal-oxide memristor crossbar via a foundry-compatible etch-down process, claiming ~99% device yield and ~26% variation in switching voltages. These metrics are presented as sufficient to program 4K-pixel gray-scale patterns with <4% average tuning error and to achieve MNIST classification with ~1% average error when importing ex-situ calculated weights.

Significance. If the experimental outcomes hold, the work advances passive crossbar neuromorphic hardware by demonstrating larger array size and analog-grade properties (yield, variation tolerance) than most prior transistor-less demonstrations. The concrete performance numbers on tuning error and classification fidelity constitute direct experimental support for the central feasibility claim.

major comments (2)
  1. [Abstract] Abstract: The MNIST result is obtained by importing pre-calculated weights rather than on-chip iterative programming. No data or procedure is shown for closed-loop conductance tuning under the reported 26% switching-voltage variation, leaving untested whether the same variation permits accurate target-reaching during repeated set/reset pulses—the step required for the claim that the crossbar supports adaptive neuromorphic circuits.
  2. [Abstract] Abstract / Results: The claim that 26% switching-voltage variation is 'sufficient for programming' with <4% tuning error rests on the gray-scale pattern experiment, yet the manuscript provides neither the raw conductance traces, the exact pulse protocol used, nor an error-propagation analysis showing how device-to-device variation maps to final conductance error. This information is load-bearing for the analog-grade assertion.
minor comments (2)
  1. [Abstract] The manuscript would benefit from a quantitative statement of the software baseline classification error for the same network size so that 'close to the software-modeled limit' can be evaluated directly.
  2. Add a brief methods paragraph or supplementary note describing the foundry-compatible etch-down parameters (temperature budget, etch chemistry) to support the reproducibility claim.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the detailed and constructive review. The comments correctly identify that the MNIST demonstration relies on ex-situ weight import and that the gray-scale tuning results would benefit from additional supporting data. We address each point below and indicate where revisions will be made to improve clarity and completeness without altering the core claims.

read point-by-point responses

  1. Referee: [Abstract] Abstract: The MNIST result is obtained by importing pre-calculated weights rather than on-chip iterative programming. No data or procedure is shown for closed-loop conductance tuning under the reported 26% switching-voltage variation, leaving untested whether the same variation permits accurate target-reaching during repeated set/reset pulses—the step required for the claim that the crossbar supports adaptive neuromorphic circuits.

    Authors: We agree that the MNIST classification experiment imports ex-situ calculated weights and does not demonstrate on-chip iterative programming of a full network. The manuscript's central experimental support for analog-grade operation under the observed variation is instead the 4K-pixel gray-scale pattern programming, which required repeated set/reset pulses to reach target conductances with <4% average error. We will revise the abstract and discussion to explicitly distinguish the ex-situ MNIST result from the on-chip tuning demonstration and to note that full closed-loop network training remains future work. revision: partial

  2. Referee: [Abstract] Abstract / Results: The claim that 26% switching-voltage variation is 'sufficient for programming' with <4% tuning error rests on the gray-scale pattern experiment, yet the manuscript provides neither the raw conductance traces, the exact pulse protocol used, nor an error-propagation analysis showing how device-to-device variation maps to final conductance error. This information is load-bearing for the analog-grade assertion.

    Authors: The gray-scale experiment applied a write-verify pulse protocol (alternating set and reset pulses with amplitude and width adjusted based on the measured switching-voltage statistics) to each device until its conductance fell within the target bin. We acknowledge that the main text summarizes the outcome without including raw traces or the precise protocol parameters. An error-propagation analysis linking the 26% voltage variation to the observed conductance error is also absent. We will add these elements—raw traces for representative devices, the exact pulse protocol, and a short error-propagation discussion—in a revised Results section or supplementary material to make the supporting evidence explicit. revision: yes

Circularity Check

0 steps flagged

No circularity: experimental fabrication and direct measurements only

full rationale

The paper reports a fabricated 64x64 passive memristor crossbar using a foundry-compatible process. All central claims (99% yield, 26% switching voltage variation, <4% tuning error on 4K patterns, ~1% MNIST import error near software limit) are direct experimental measurements of fabricated devices and ex-situ weight import. No equations, derivations, fitted parameters renamed as predictions, or self-citation chains appear in the provided text. The derivation chain is empty; results do not reduce to inputs by construction.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

The central claim rests on experimental fabrication outcomes. No free parameters are introduced. The process relies on standard semiconductor fabrication assumptions rather than new postulates.

axioms (1)
  • domain assumption A foundry-compatible fabrication process featuring etch-down patterning and low-temperature budget is conducive to vertical integration and can produce functional passive crossbars.
    Invoked in the abstract to justify the process choice and explain the achieved yield and integration potential.

pith-pipeline@v0.9.0 · 5760 in / 1176 out tokens · 44709 ms · 2026-05-25T14:23:01.710371+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

What do these tags mean?
matches
The paper's claim is directly supported by a theorem in the formal canon.
supports
The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends
The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses
The paper appears to rely on the theorem as machinery.
contradicts
The paper's claim conflicts with a theorem or certificate in the canon.
unclear
Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Reference graph

Works this paper leans on

50 extracted references · 50 canonical work pages · 1 internal anchor

  1. [1]

    & Cauwenberghs, G

    Chakrabartty, S. & Cauwenberghs, G. Sub-microwatt analog VLSI trainable pattern classifier. IEEE J.Solid-State Circuits 42, 1169-1179 (2007)

    work page 2007

  2. [2]

    & Hasler, J

    Ramakrishnan, S. & Hasler, J. Vector-matrix multiply and winner-take-all as an analog classifier. IEEE Trans. Very Large Scale Integr. Syst. 22, 353–361 (2014)

    work page 2014

  3. [3]

    K., & Strukov, D

    Merrikh Bayat, F., Guo, X., Klachko, M., Prezioso, M., Likharev, K. K., & Strukov, D. B. High- performance mixed-signal neurocomputing with nanoscale floating-gate memory cells. IEEE Trans. Neural Netw. Learn. Syst. 29, 4782–4790 (2018)

    work page 2018

  4. [4]

    B., Kuzum, D., Jeyasingh, R., Kim, S., BrightSky, M., Lam, C., Wong, H

    Eryilmaz, S. B., Kuzum, D., Jeyasingh, R., Kim, S., BrightSky, M., Lam, C., Wong, H. S. P. Brain-like associative learning using a nanoscale non-volatile phase change synaptic device array. Front. Neurosci. 8, 205 (2014)

    work page 2014

  5. [5]

    Burr, G. W. et al. Experimental demonstration and tolerancing of a large-scale neural network (165 000 synapses) using phase-change memory as the synaptic weight element. IEEE Trans. Electron Devices 62, 3498-3507 (2015)

    work page 2015

  6. [6]

    L., Nandakumar, S

    Boybat, I., Gallo, M. L., Nandakumar, S. R., Moraitis, T., Parnell, T., Tuma, T., Rajendran, B., Leblebici, Y., Sebastian, A. & Eleftheriou, E. Neuromorphic computing with multi-memristive synapses. Nat. Commun. 9, 2514 (2018)

    work page 2018

  7. [7]

    & Ueda, M

    Kaneko, Y., Nishitani, Y. & Ueda, M. Ferroelectric artificial synapses for recognition of a multishaded image. IEEE Trans. Electron Devices 61, 2827–2833 (2014)

    work page 2014

  8. [8]

    Boyn, S. et al. Learning through ferroelectric domain dynamics in solid-state synapses. Nat. Commun. 8, 14736 (2017)

    work page 2017

  9. [9]

    Romera, M. et al. Vowel recognition with four coupled spin-torque nano-oscillators. Nature 563, 230-234 (2018)

    work page 2018

  10. [10]

    H., Kim, K

    Jo, S. H., Kim, K. H. & Lu, W. High-density crossbar arrays based on a Si memristive system. Nano Lett. 9, 870-874 (2009)

    work page 2009

  11. [11]

    Milano, G. et al. Self-limited single nanowire systems combining all-in-one memristive and neuromorphic functionalities. Nat. Commun. 9, 5151 (2018)

    work page 2018

  12. [12]

    Fuller, E. J. et al. Parallel programming of an ionic floating-gate memory array for scalable neuromorphic computing. Science 364, 570-574 (2019)

    work page 2019

  13. [13]

    Indiveri, G. et al. Integration of nanoscale memristor synapses in neuromorphic computing architectures. Nanotechnology 24, 384010 (2013)

    work page 2013

  14. [14]

    Prezioso, M. et al. Training and operation of an integrated neuromorphic network based on metal-oxide memristors. Nature 521, 61-64 (2015)

    work page 2015

  15. [15]

    Prezioso, M. et al. Spike-timing-dependent plasticity learning of coincidence detection with passively integrated memristive circuits. Nature Commun. 9, 5311 (2018)

    work page 2018

  16. [16]

    Ambrogio, S. et al. Neuromorphic learning and recognition with one-transistor-one-resistor synapses and bistable metal oxide RRAM. IEEE Trans. Electron Devices 63, 1508-1515 (2016)

    work page 2016

  17. [17]

    Sheridan, P. M. et al. Sparse coding with memristor networks. Nat. Nanotechnol. 12, 784-789 (2017)

    work page 2017

  18. [18]

    Yao, P. et al. Face classification using electronic synapses. Nat. Commun. 8, 15199 (2017)

    work page 2017

  19. [19]

    Yu, S. et al. Binary neural network with 16 Mb RRAM macro chip for classification and online training. in Proc. IEEE International Electron Device Meeting (IEDM), 16.2.1-16.2.4 (IEEE, 2016)

    work page 2016

  20. [20]

    Hu, M. et al. Memristor‐based analog computation and neural network classification with a dot product engine. Adv. Mater. 30, 1705914 (2018)

    work page 2018

  21. [21]

    Merrikh Bayat, F. et al. Implementation of multilayer perceptron network with highly uniform passive memristive crossbar circuits. Nat. Commun. 9, 2331 (2018)

    work page 2018

  22. [22]

    J., Strukov, D

    Yang, J. J., Strukov, D. B. & Stewart, D. Memristive devices for computing. Nat. Nanotechnol. 8, 13-14 (2013)

    work page 2013

  23. [23]

    et al., Recommended methods to study resistive switching devices, Adv

    Lanza, M. et al., Recommended methods to study resistive switching devices, Adv. Electron. Mater. 5, 1800143 (2019). H. Kim et al., Analog-Grade Passive Memristive Circuit, June 2019 10

    work page 2019

  24. [24]

    Redox‐based resistive switching memories (ReRAMs): Electrochemical systems at the atomic scale

    Valov, I. Redox‐based resistive switching memories (ReRAMs): Electrochemical systems at the atomic scale. ChemElectroChem 1, 26-36 (2014)

    work page 2014

  25. [25]

    A review of emerging non-volatile memory (NVM) technologies and applications

    Chen, A. A review of emerging non-volatile memory (NVM) technologies and applications. Solid-State Electron. 125, 25-38 (2015)

    work page 2015

  26. [26]

    Park, K. et al. Three-dimensional 128 Gb MLC vertical NAND Flash memory with 24-WL stacked layers and 50 MB/s high-speed programming. IEEE J. Solid-State Circuits 50, 204-213 (2015)

    work page 2015

  27. [27]

    Shafiee, A. et al. ISAAC: A convolutional neural network accelerator with in-situ analog arithmetic in crossbars. ACM SIGARCH Comput. Architect. News 44, 14-26 (2016)

    work page 2016

  28. [28]

    Bavandpour, M. et al. Mixed-signal neuromorphic inference accelerators: Recent results and future prospects. in Proc. IEEE International Electron Device Meeting (IEDM), 20.4.1-20.4.4 (IEEE, 2018)

    work page 2018

  29. [29]

    Mahmoodi, M. R. & Strukov, D. B. An ultra-low energy internally analog, externally digital vector-matrix multiplier circuit based on NOR flash memory technology. in Proc. ACM/ESDA/IEEE Design Automation Conference (DAC) (2018)

    work page 2018

  30. [30]

    Amodei, D. et al. Deep speech 2: End-to-end speech recognition in English and Mandarin. in Proc. International Conference on Machine Learning (ICML), 173-182 (2016)

    work page 2016

  31. [31]

    Johnson, M. et al. Google’s multilingual neural machine translation system: Enabling zero-shot translation. Trans. Asso. Comput. Linguist. 5, 339-351 (2017)

    work page 2017

  32. [32]

    Shazeer1, N. et al. Outrageously large neural networks: The sparsely-gated mixture-of-experts layer. Preprint at https://arxiv.org/abs/1701.06538 (2017)

    work page internal anchor Pith review Pith/arXiv arXiv 2017

  33. [33]

    The Cerebral Cortex, Harvard U

    Mountcastle, V. The Cerebral Cortex, Harvard U. Press, 1998

    work page 1998

  34. [34]

    Chou, C. et al. An N40 256Kx44 embedded RRAM macro with SL-precharge SA and low-voltage current limiter to improve read and write performance. in Solid-State Circuits Conference Digest of Technical papers (ISSCC), 478-479, (IEEE, 2018)

    work page 2018

  35. [35]

    Burr, G. W. et al. Neuromorphic computing using non-volatile memory. Adv. Phys.: X. 2, 89-124 (2017)

    work page 2017

  36. [36]

    & Yang, J

    Xia, Q. & Yang, J. J. Memristive crossbar arrays for brain-inspired computing. Nat. Mater. 18, 309-323 (2019)

    work page 2019

  37. [37]

    & Alibart, F

    Rajendran, B. & Alibart, F. Neuromorphic computing based on emerging memory technologies. IEEE J. Emerg. Sel. Top. Circuits Syst. 6, 198-211 (2016)

    work page 2016

  38. [38]

    Neuro-inspired computing with emerging nonvolatile memories

    Yu, S. Neuro-inspired computing with emerging nonvolatile memories. Proc. IEEE 106, 260-285 (2018)

    work page 2018

  39. [39]

    Strukov, D. B. Tightening grip. Nat. Mater. 17, 293-295 (2018)

    work page 2018

  40. [40]

    Kau, D. et al. A stackable cross point phase change memory. in Proc. IEEE International Electron Device Meeting (IEDM), 27.1.1- 27.1.4 (IEEE, 2009)

    work page 2009

  41. [41]

    Choi, S. et al. SiGe epitaxial memory for neuromorphic computing with reproducible high performance based on engineered dislocations. Nat. Mater. 17, 335-341 (2018)

    work page 2018

  42. [42]

    Goswami, S. et al. Robust resistive memory devices using solution-processable metal-coordinated azo aromatics. Nat. Mater. 16, 1216–1224 (2017)

    work page 2017

  43. [43]

    Kim, K. H. et al. A functional hybrid memristor crossbar-array/CMOS system for data storage and neuromorphic applications. Nano Lett. 12, 389–395 (2012)

    work page 2012

  44. [44]

    Yang, J. J. et al. Memristive switching mechanism for metal/oxide/metal nanodevices. Nat. Nanotechnol. 3, 429-433 (2008)

    work page 2008

  45. [45]

    Hoskins, B. D. & Strukov, D. B. Maximizing stoichiometry control in reactive sputter deposition of TiO2. J. Vac. Sci. Technol. A 35, 20606 (2017)

    work page 2017

  46. [46]

    Kaneko, A. et al. Sidewall transfer process and selective gate sidewall spacer formation technology for sub-15nm FinFET with elevated source/drain extension. in Proc. IEEE International Electron Device Meeting (IEDM), 844-847. (IEEE, 2005)

    work page 2005

  47. [47]

    Adam, G. C. et al. 3-D memristor crossbars for analog and neuromorphic computing applications. IEEE Trans. Electron Devices 64, 312-318 (2017). H. Kim et al., Analog-Grade Passive Memristive Circuit, June 2019 11

    work page 2017

  48. [48]

    & Strukov, D

    Alibart, F., Gao, L., Hoskins, B. & Strukov, D. B. High-precision tuning of state for memristive devices by adaptable variation-tolerant algorithm. Nanotechnology 23, 075201 (2012)

    work page 2012

  49. [49]

    Suh, K. D. et al. A 3.3 V 32 Mb NAND flash memory with incremental step pulse programming scheme. IEEE J. Solid-State Circuits 30, 1149-1156 (1995)

    work page 1995

  50. [50]

    The MNIST database of handwritten digits

    “The MNIST database of handwritten digits”, online resource, available at http://yann.lecun.com/exdb/mnist/ H. Kim et al., Analog-Grade Passive Memristive Circuit, June 2019 12 Figure 1. Memristive crossbar array. (a) Scanning electron microscope image of the fabricated 64×64 memristor crossbar array. Top, bottom left, and bottom right insets show, corres...

    work page 2019