Logo

Novel memristor design clears critical impediments for future AI chips
(Nanowerk Spotlight) Researchers are pursuing memristors – resistive memory devices with properties similar to neurons – as a means to develop energy-efficient hardware that can accelerate artificial intelligence. However, multiple obstacles have impeded memristors’ path to widespread adoption, from insufficient current density to sneaking currents that disrupt large-scale integration. Now, researchers at Lund University report a breakthrough memristor built from ferroelectric hafnium oxide that clears these hurdles through strong nonlinear current-voltage characteristics combined with ultra-low conductance operation.
The findings have been published in Advanced Intelligent Systems ("Ferroelectric Tunnel Junction Memristors for In-Memory Computing Accelerators").
crossbar implementation of FTJ memristors where the ferroelectric (teal) is sandwiched between the top (red) and bottom (blue) electrodes
a) True crossbar implementation of FTJ memristors where the ferroelectric (teal) is sandwiched between the top (red) and bottom (blue) electrodes. b) Implementation of multiply and accumulate operations in a crossbar arrangement using analog temporal encoding. The input xi is encoded in the pulse length tn using a constant amplitude V0. The currents I are summed up though each bit-line of the array where the magnitude of the current is dependent on the programmed memristor conductance G. The current is then integrated to get the charge Q. c) Fabrication process of FTJ devices. (I) Deposition of the TiN bottom electrode on Si/SiO2 substrate using RF Sputtering. (II) ALD growth of amorphous HfxZr1-xO2 followed by (III) W crystallization electrode deposition and RTP at 535 °C. (IV)–(VI) metal replacement process and W top electrode deposition and patterning through UV-lithography and lift-off process. (Reprinted with permission from Wiley-VCH Verlag)
Since the dawn of the artificial intelligence (AI) revolution, innovators have struggled with the fundamental limitations of traditional computing architectures for training and running data-hungry neural networks. Called the “von Neumann bottleneck,” this constraint arises from the shuttling of data back and forth between a computer’s processor and memory. DARPA launched the Memristor Discovery and Development program in 2008 to explore resistors with memory that could collapse the distinction between memory and logic. Regarded as electronic analogs to biological synapses, memristors offered a path to the brain-inspired computing paradigm known as neuromorphic engineering.
However, after years of halting progress, researchers have yet to develop memristors that check all the boxes for integration into dense, ultra-efficient hardware accelerators for AI. Previous designs have fallen short due to insufficient current density through the ultra-thin tunnel barrier, sneak currents that degrade readout accuracy in large crossbar arrays, or failure to meet key memristor performance targets liked symmetric weight update. These lingering deficiencies have blocked memristors from escaping niche research applications into mainstream computing.
The new study demonstrates ferroelectric tunnel junction (FTJ) memristors built on a hafnium-zirconium oxide switching layer that appears poised to overcome these prior limitations. Led by Dr. Robin Athle and Dr. Mattias Borg from Lund University, the researchers optimized their FTJ memristors to achieve high tunneling current density exceeding 3 A/m2, more than 60 incremental conductance states, wide dynamic range between on and off states, and robust data retention over 100 seconds.
Unlike previous attempts with hafnium oxide FTJs limited by low current density during readout, the team’s devices benefit from an ultra-thin, sub-5 nm ferroelectric film that greatly enhances charge transport while preserving strong polarization responsiveness. This permits their memristors to be aggressively scaled down to dimensions compatible with advanced CMOS nodes without deteriorating performance. In fact, read energy could plausibly reach levels around 30 femtojoules per bit with further size reductions.
But beyond improving a longstanding current density limitation for hafnium oxide FTJs, Athle and Borg’s memristors excel in areas that matter most for hardware-based neural network training. Using an amplitude modulation scheme to incrementally program device conductance, they achieve symmetrical potentiation and depression behavior along with appreciable linearity. This predictability allows their FTJs to delineate over 60 distinct, progressive conductance levels that enable precise tuning of synaptic weight analogs during online learning. And with device nonlinearity exceeding 1000, the researchers’ FTJs intrinsically minimize disruptive sneak currents so that selector elements become unnecessary in large crossbar arrays.
The team next set out to validate whether their FTJ memristors’ properties would translate to accurate neuromorphic computing applications. They partnered with collaborators to simulate a neural network implemented using FTJ crossbars on a modified dataset of handwritten digits. Despite flaws from device variability and nonlinear conductance modulation, their simulated array attained 92% accuracy in classifying the dataset - on par with other cutting-edge memristor technologies.
But equally noteworthy, the researchers conduit extensive simulations that suggest their FTJs’ extremely low conductance conveys game-changing resilience to parasitic effects that have beleaguered densely-integrated crossbar arrays. Analysis indicates their memristor crossbars could mitigate debilitating IR drops far more effectively than existing resistive memory alternatives, potentially allowing for 150x larger arrays. This size versatility arises organically from the FTJs’ intrinsic device physics rather than extra circuitry.
“Overall, this study highlights the potential of using hafnium-zirconium oxide based FTJs as memristive elements in future neuromorphic applications to accelerate neural network training and inference,” concluded study co-author Dr. Mattias Borg.
With performance indicators consistently matching or overriding previous memristors along with a high tolerance to integration pitfalls, Athle and Borg’s FTJ devices seize tremendous promise. Their compelling results suggest that after years of nominal progress, the memristor drought may soon transform into an oasis, priming these long-hyped devices to finally permeate AI computing architectures and deliver energy efficiency and speed gains far surpassing status quo silicon.
If robustness persists as research and development continues, this breakthrough could open the floodgates to an array of accelerated deep learning applications - from real-time video analytics to perceptive autonomous robots - previously deemed too computationally unwieldy for edge devices.
Michael Berger By – Michael is author of three books by the Royal Society of Chemistry:
Nano-Society: Pushing the Boundaries of Technology,
Nanotechnology: The Future is Tiny, and
Nanoengineering: The Skills and Tools Making Technology Invisible
Copyright © Nanowerk LLC
 

Become a Spotlight guest author! Join our large and growing group of guest contributors. Have you just published a scientific paper or have other exciting developments to share with the nanotechnology community? Here is how to publish on nanowerk.com.