Device and materials requirements for neuromorphic computing

Abstract

Energy efficient hardware implementation of artificial neural network is challenging due the 'memory-wall' bottleneck. Neuromorphic computing promises to address this challenge by eliminating data movement to and from off-chip memory devices. Emerging non-volatile memory (NVM) devices that exhibit gradual changes in resistivity are a key enabler of in-memory computing-a type of neuromorphic computing. In this paper, we present a review of some of the NVM devices (RRAM, CBRAM, PCM) commonly used in neuromorphic application. The review focuses on the trade-off between device parameters such as retention, endurance, device-to-device variation, speed and resistance levels, and the interplay with target applications. This work aims at providing guidance for finding the optimized resistive memory devices material stack suitable for neuromorphic application.

Bridging Biological and Artificial Neural Networks with Emerging Neuromorphic Devices: Fundamentals, Progress, and Challenges

“…[74,133,134] Depending on the applications and algorithms, different switching characteristics may be emphasized, such as switching symmetry, linearity, number of conductance states. From electronics point of view, the key feature of biological synapses is tunable weight, which can be conveniently translated to the conductance (or resistance) change in electronic devices.…”

Silver Nanofilament Formation Dynamics in a Polymer‐Ionic Liquid Thin Film by Direct Write

“…Because the current at 1.8 V changes gradually with time after the current reaches i n , it suggests the possibility of achieving multiple resistance states-a requirement for neuromorphic computing [9,23,32] -during a single nanofilament formation under constant formation bias. To test whether or not the PEGDA/IL system can accommodate such a requirement, we set the system compliance to 100 nA and selected eight nanofilament "programming currents" ranging from 15 to 50 nA with a step of 5 nA.…”

“…The conductance initially decreases on the timescale of ms, Figure 5. [23,32] In most cases, repeated programming pulses are used to controllably adjust resistance, [23,32] but Figure 6 strongly suggests that multiple resistance states may be achieved via the competition between the EDL formation and Ag redox reaction-thereby introducing a new mechanism to tune resistance. a) Schematic of abrupt (left) versus gradual (right) growth.…”

Memristive Devices and Networks for Brain‐Inspired Computing

“…Nevertheless, offline learning applications places a high requirement on retention (>10 years) to allow reliable inference after the synaptic weights in the arrays are programmed. However, such retention requirement can be relaxed in online learning, where the weights are updated frequently …”

“…In particular, a dynamic range of 100, weight precision of 6 bit and low energy consumption that is comparable with biological synapses (≈10 fJ) might be desirable for many applications . It is worthwhile pointing out that other requirements on device performance may strongly depend on detailed applications, and Table only suggests typical values. For example, an endurance of 10 5 may be required to train state‐of‐the‐art neural networks in an online fashion, but it should be noted that the device conductance only needs to be modified by an incremental amount instead of being switched across the whole dynamic range in each iteration, making the requirement for endurance a bit relaxed.…”