Source-linked AI summary

Magnetic domain wall based synaptic and activation function generator for neuromorphic accelerators

Saima A Siddiqui, Sumit Dutta, Astera Tang, Luqiao Liu, Caroline A Ross, Marc A Baldo

arXiv:1909.06026v1cond-mat.dis-nncond-mat.mes-hallcs.ET

TL;DR

Domain-wall devices offer an approach to neuromorphic computation that can implement linear synaptic weights and programmable nonlinear activation functions. The paper demonstrates discrete and quasi-analog devices, with prototype operation using 8 ns pulses and energy use of 1–16 pJ.

  • Problem

    Nonvolatile devices can implement neural-network linear operations, but their material- and structure-specific conductance nonlinearities are difficult to customize for linear synaptic functions or programmable nonlinear activations.

    Source: Discussion · p. 10
  • Method

    The paper uses current-driven domain-wall motion beneath MTJs to generate discrete synaptic weights and nonlinear activations, and demonstrates quasi-analog wall motion in a CoTb wire.

    Source: Discretized synaptic weight generation · p. 6, Discretized activation function… · p. 8, Analog implementation of activation… · p. 9
  • Results

    Prototype devices operated with 8 ns current pulses and consumed 1–16 pJ for switching; scaled-device energy is predicted to approach that of biological synapses.

    Source: Discussion · pp. 10–11
  • Takeaways & Limitations

    The demonstrated linear and nonlinear behaviors show that three-terminal MTJs can be customized for multiple roles in a neuromorphic accelerator.

    Source: Discussion · p. 9

    The patterned reference-layer stray field can trap domain walls, so the demonstrated devices require a compensating external bias field.

    Source: Magnetic tunnel junctions… · p. 5

Abstract

from arXiv · show

Magnetic domain walls are information tokens in both logic and memory devices, and hold particular interest in applications such as neuromorphic accelerators that combine logic in memory. Here, we show that devices based on the electrical manipulation of magnetic domain walls are capable of implementing linear, as well as programmable nonlinear, functions. Unlike other approaches, domain-wall-based devices are ideal for application to both synaptic weight generators and thresholding in deep neural networks. Prototype micrometer-size devices operate with 8 ns current pulses and the energy consumption required for weight modulation is < 16 pJ. Both speed and energy consumption compare favorably to other synaptic nonvolatile devices, with the expected energy dissipation for scaled 20 nm devices close to that of biological neurons.

Magnetic tunnel junctions characterizations

Perpendicular MTJs read the domain orientation electrically, while reference-layer stray fields affect switching and require compensation. Current pulses move domain walls beneath the junctions to switch their resistance state.

Magnetic tunnel junctions characterizations

  • The perpendicular MTJ converts magnetic orientation into resistance through the relative magnetization alignment of CoFeB layers separated by MgO.The bottom CoFeB is the programmable free layer; the top CoFeB and synthetic antiferromagnet form the reference layer.Source: Magnetic tunnel junctions… · p. 4
  • A 1 μm × 3 μm MTJ has approximately 44% tunnel magnetoresistance, and its free-layer minor loop is offset from zero field by the reference layer’s stray field.Source: Magnetic tunnel junctions… · pp. 4–5
  • The patterned synthetic antiferromagnetic reference layer creates a stray field that can trap domain walls; device operation here uses a compensating external bias field.The authors also describe alternative designs that eliminate the stray field.Source: Magnetic tunnel junctions… · p. 5
  • A magnetic-field range creates domain walls at reference-layer edges, consistent with the differing switching fields inside and outside the MTJ region.Source: Magnetic tunnel junctions… · p. 5
  • Bias-field-assisted current pulses translate a domain wall under the MTJ to switch it; lower bias fields require higher switching current density, and a larger bias field resets the device.The bias field both nucleates the walls and counters reference-layer stray fields.Source: Magnetic tunnel junctions… · pp. 5–6

Discretized synaptic weight generation

A wire with linearly varying width sets a linear current-density profile across parallel MTJs, which switch stepwise to produce a linear resistance-versus-current response.

Discretized synaptic weight generation

  • The device uses parallel 200 nm × 400 nm MTJs on a CoFeB free layer whose width, along with the underlying Ta, varies linearly to change current density along the wire.A suitable bias field supports a discrete magnetic domain at each MTJ.Source: Discretized synaptic weight generation · p. 6
  • Resistance changes linearly with applied current, providing a direct mapping from synaptic weights in the algorithms to device conductance.Source: Discretized synaptic weight generation · p. 7
  • Eight resistance values correspond to consecutive switching of seven MTJs; two MTJs on the wire’s widest part did not switch because their required current exceeded the narrowest part’s breakdown current.Source: Discretized synaptic weight generation · p. 7

Discretized activation function generation

The activation generator uses parallel MTJs with varying areas and wire width to approximate a sigmoid; measured switching follows the designed response for seven MTJs.

Discretized activation function generation

  • The device approximates a sigmoid by varying MTJ areas, with larger devices at the activation curve’s steepest portion, and linearly increasing wire width to set regularly spaced switching currents.Source: Discretized activation function… · p. 7
  • Averaged over 75 cycles, the seven switching MTJs follow the designed response and target shifted sigmoid as input current pulses ramp from zero to 620 μA at 200 Oe.The remaining two devices switch at lower fields.Source: Discretized activation function… · p. 8

Analog implementation of activation function

The analog design replaces discrete domain storage with a quasi-continuous domain-wall position and shapes the current-carrying wire to produce sigmoid-like travel distances.

Analog implementation of activation function

  • In an analog device, domain-wall position is quasi-continuous, with spatial resolution fundamentally limited by the fabrication process’s patterning correlation length.Source: Analog implementation of activation… · pp. 8–9
  • A Pt/CoTb/SiNx device without a reference layer allows a domain wall to reside along the wire, and measured travel distances follow the designed sigmoid after 300 ns current pulses.Walls are initialized at position A for each measurement.Source: Analog implementation of activation… · p. 9
  • The analog demonstration detects wall motion magneto-optically; a continuous MTJ reference layer could be designed to provide electrical detection of domain-wall positions.Source: Analog implementation of activation… · p. 9

Discussion

The discussion presents three-terminal MTJs as customizable devices for linear and nonlinear neuromorphic functions, with low switching energy and silicon-process compatibility. It also identifies design options and practical constraints for scaling and array use.

Discussion

  • The demonstrated linear and nonlinear behaviors show that three-terminal MTJs can serve multiple neuromorphic-accelerator roles.The devices are also described as compatible with traditional silicon processing.Source: Discussion · pp. 9–10
  • Individual MTJs can be selectively programmed after fabrication by initializing magnetic domains, and engineered wire widths and junction areas can reproduce linear or nonlinear characteristics without continuous free layers.The proposed design can also adjust MgO barrier thickness to reach mega-ohm synaptic-generator resistance for cross-point arrays.Source: Discussion · pp. 10–11
  • Prototype MTJs use 8-ns pulses with 1–16 pJ switching energy; scaled 20-nm devices are projected at ~18–36 fJ per synaptic event, assuming shorter retention is acceptable.The authors expect scaled-device energy to be lower than other nonvolatile devices and within one order of magnitude of biological synapses.Source: Discussion · p. 10
  • The authors summarize demonstrations of a domain-wall linear synaptic-weight generator and nonlinear activation-function generator, with proposed silicon-foundry compatibility.They characterize the technology as a candidate for neuromorphic accelerators based on its computation-and-memory versatility, latency, and energy use.Source: Discussion · p. 11

Sample preparation

The MTJ films are sputtered, annealed, and patterned through lithography and ion milling; CoTb samples are separately deposited and patterned for magneto-optic measurements.

Sample preparation

  • MTJ films are sputtered onto thermally oxidized silicon, then annealed in nitrogen at 250 °C for 15 minutes.The stack uses DC magnetron sputtering for thin films and RF sputtering for the MgO barrier.Source: Sample preparation · p. 12
  • Three electron-beam lithography and two Ar+ ion-milling steps define the MTJ free-layer wires, junctions, and contacts, with SiO2 separating top and bottom contacts.Ti/Au contacts are evaporated and lifted off in the final patterning step.Source: Sample preparation · p. 12
  • CoTb samples are co-sputtered, capped with SiNx, and patterned in separate e-beam steps for the bottom Pt wire and constant-width CoTb/SiNx wire.Processing temperature remains below 100 °C.Source: Sample preparation · p. 12

Electrical measurement

Electrical characterization uses four-contact anomalous Hall measurements for wires and two-contact magnetoresistance measurements for tunnel junctions. Nanosecond pulses drive domain-wall motion, while Kerr imaging characterizes CoTb motion.

Electrical measurement

  • Wires are measured by four-contact anomalous Hall resistance, while tunnel-junction magnetoresistance is measured with two contacts.Quasi-DC measurements use a Keithley 2400 source meter; the tunnel-junction measurement current is 100 nA.Source: Electrical measurement · p. 13
  • Current-induced domain-wall motion is driven by nanosecond pulses through CoFeB wires, with tunnel magnetoresistance measured using a DC probe.The setup uses an RF probe and either an Agilent 8130A or 8114A pulse generator.Source: Electrical measurement · p. 13
  • For CoTb magneto-optic Kerr measurements, nanosecond pulses are applied and images are captured before and after each pulse.The devices are mounted on an RF chip carrier with a 5-GHz cutoff and imaged under in-plane and out-of-plane magnetic fields.Source: Electrical measurement · p. 13

Corresponding authors

Correspondence is directed to Saima A Siddiqui or Marc A Baldo.

Corresponding authors

Loading 1909.06026v1…