Compute in‐Memory with Non‐Volatile Elements for Neural Networks: A Review from a Co‐Design Perspective

Abstract

Deep learning has become ubiquitous, touching daily lives across the globe. Today, traditional computer architectures are stressed to their limits in efficiently executing the growing complexity of data and models. Compute‐in‐memory (CIM) can potentially play an important role in developing efficient hardware solutions that reduce data movement from compute‐unit to memory, known as the von Neumann bottleneck. At its heart is a cross‐bar architecture with nodal non‐volatile‐memory elements that performs an analog multiply‐and‐accumulate operation, enabling the matrix‐vector‐multiplications repeatedly used in all neural network workloads. The memory materials can significantly influence final system‐level characteristics and chip performance, including speed, power, and classification accuracy. With an over‐arching co‐design viewpoint, this review assesses the use of cross‐bar based CIM for neural networks, connecting the material properties and the associated design constraints and demands to application, architecture, and performance. Both digital and analog memory are considered, assessing the status for training and inference, and providing metrics for the collective set of properties non‐volatile memory materials will need to demonstrate for a successful CIM technology.

Future Directions Workshop: Materials, Processes, and R&D Challenges in Microelectronics

“…However, these will not be energy efficient enough for the large data-intensive workloads of the future and will require new approaches to matrix multiplication hardware. One of the options is "compute-inmemory" by using a Kirchoff's Law based analog step to carry out a vector dot product that nearly eliminates the shuttling of data between compute and memory through a crossbar array architecture (Haensch et al, 2022;W. Wan et al, 2022).…”

Impact of edge defects on the synaptic characteristic of a ferromagnetic domain-wall device and on on-chip learning

“…Among the various non Von Neumann/neuromorphic computing schemes recently proposed and implemented, crossbar arrays based on NVM devices as synapses can be deemed to be the most popular. These crossbar arrays carry out analog computing (both inference and training of neural networks (NN)) in analog memory systems (in-memory computing) very efficiently, both in terms of speed and energy [1][2][3][4][5]. During forward computation/inference, the crossbar array carries out vector-matrix multiplication (VMM) between the input vector and the synaptic weight matrix very fast by making use of Ohm's Law and Kirchoff 's Current Law.…”

The Dynamics of Oxygen Ion Exchange in Epitaxial Strontium Cobaltite Bilayers

“…Interestingly, Mou et al recently showed that the topotactic transition in SrCoO x could be highly beneficial for memristive applications since the resistive switching process is likely to be better controlled than in most filamentary systems, [18] and other reports have similarly demonstrated interest in this and related systems for memristive applications. [20,23,24] Here, the ordered vacancy channels allow a pre-defined path for oxygen migration without structural breakdown. From a thermodynamic perspective, the energy barrier for switching between the two phases is relatively small -smaller than those of similar systems such as strontium manganite (SrMnO 2.5 and SrMnO 3 − 𝛿 ) [11] -and the application of a bias to a BM-SCO/PV-SCO heterostructure could be used to change the overall resistance of a stack, paving the way for controllable and reversible synaptic memories.…”