Atomically Tunable Resistive Random Access Memory: Prospective for Neuromorphic Computing

Category Artificial Intelligence

tldr #

Continuous device downsizing and circuit complexity have motivated the need for atomically-tunable memristor devices.Researchers have created spiking neural nets with thousands of memristors (that mimic how neurons operate) by using atomically thin layer of material. The CMOS transistors provide an outstanding control over the current across the memristors and demonstrate in-memory computation by constructing logic gates. Developments in resistive random access memory (RRAM) have enabled researchers to mimic biological neurons and synapses using memristors, which serves as a trainable synapses in neural networks. Advancements towards atomically-tunable memristors will allow for a new generation of neuromorphic computers.


content #

Researchers have created spiking neural nets with thousands of memristors (that mimic how neurons operate) by using atomically thin layer of material. Spiking neural networks should be better than regular neural networks where there is the need for intelligence to react to real-time inputs. Regular neural networks are better for more static information.

Researchers transfer a sheet of multilayer hexagonal boron nitride (h-BN) onto the back-end-of-line (BEOL) interconnections of silicon microchips containing CMOS transistors of the 180 nanometer node, and finalize the circuits by patterning the top electrodes and interconnections. The CMOS transistors provide outstanding control over the currents across the h-BN memristors, which allows us to achieve endurances of ~5 million cycles in memristors as small as ~0.053 square micron. They demonstrate in-memory computation by constructing logic gates, and measure spike-timing dependent plasticity (STDP) signals that are suitable for the implementation of spiking neural networks (SNN). The high performance and the relatively-high technology readiness level achieved represent a significant advance towards the integration of 2D materials in microelectronic products and memristive applications.

The combination of CMOS transistors and memristors are used to achieve a unified platform for powering spiking neural networks.

Abstract .

Continuous device downsizing and circuit complexity have motivated atomic-scale tuning of memristors. Herein, we report atomically tunable Pd/M1/M2/Al ultrathin (<2.5 nm M1/M2 bilayer oxide thickness) memristors using in vacuo atomic layer deposition by controlled insertion of MgO atomic layers into pristine Al2O3 atomic layer stacks guided by theory predicted Fermi energy lowering leading to a higher high state resistance (HRS) and a reduction of oxygen vacancy formation energy. Excitingly, memristors with HRS and on/off ratio increasing exponentially with M1/M2 thickness in the range 1.2–2.4 nm have been obtained, illustrating tunneling mechanism and tunable on/off ratio in the range of 10–10^4. Further dynamic tunability of on/off ratio by electric field is possible by designing of the atomic M2 layer and M1/M2 interface. This result probes ways in the design of memristors with atomically tunable performance parameters.

CMOS transistors provide an outstanding control over the current across the memristors.

Introduction .

Interest in neuromorphic computing has been steadily increasing in recent years, due to its potential to circumvent the Von Neumann bottleneck which arises from the extra energy and time required to transport data between memory and processor units during computation. Spearheaded by improvements in resistive random access memory (RRAM), researchers have attempted to mimic biological neuron and synapse operation by using artificial versions of these elements known as memristors. In memristors, the operations of both non-volatile memory storage and low-power computing are integrated in one device, taking advantage of the computing strengths of a neural system, namely pattern recognition and unstructured data sorting. RRAM may be used to implement multiple resistance states and thus serve as trainable synapses in neural networks. However, as the field of memristors, RRAM, and neuromorphic computing evolves, the ability to tune the memristive resistance, switching speed, cycling endurance, among other performance criteria, at the atomic scale will be important as device architectures evolve a new generation of neuromorphic computers. Developing memristors that perform reliably and exhibit adjustable characteristics has been a primary challenge for the computing community. In this article, we will discuss the potential of atomically-tunable memristor devices, the effect they have on the device development process and their readiness for integration into existing and future neuromorphic applications.

Neuromorphic computing is a form of artificial intelligence that mimics the functioning of the human brain.

hashtags #
worddensity #

Share