Traditional von Neumann computing systems involve separate processing and memory units. However, data movement is costly in terms of time and energy and this problem is aggravated by the recent explosive growth in highly data-centric applications related to artificial intelligence. This calls for a radical departure from the traditional systems and one such non-von Neumann computational approach is in-memory computing. Hereby certain computational tasks are performed in place in the memory itself by exploiting the physical attributes of the memory devices. Both charge-based and resistance-based memory devices are being explored for in-memory computing. In this Review, we provide a broad overview of the key computational primitives enabled by these memory devices as well as their applications spanning scientific computing, signal processing, optimization, machine learning, deep learning and stochastic computing. This Review provides an overview of memory devices and the key computational primitives for in-memory computing, and examines the possibilities of applying this computing approach to a wide range of applications.

Memory devices and applications for in-memory computing

A unified algorithm for elementary functions

High-dimensionality matrix-vector multiplication (MVM) is a dominant kernel in signal-processing and machine-learning computations that are being deployed in a range of energy- and throughput-constrained applications. In-memory computing (IMC) exploits the structural alignment between a dense 2D array of bit cells and the dataflow in MVM, enabling opportunities to address computational energy and throughput. Recent prototypes have demonstrated the potential for 10 t benefits in both metrics. However, fitting computation within an array of constrained bit-cell circuits imposes a number of challenges, including the need for analog computation, efficient interfacing with conventional digital accelerators (enabling the required programmability), and efficient virtualization of the hardware to map software. This article provides an overview of the fundamentals of IMC to better explain these challenges and then identifies promising paths forward among the wide range of emerging research.

In-Memory Computing: Advances and prospects

Large-scale matrix-vector multiplications, which dominate in deep neural networks (DNNs), are limited by data movement in modern VLSI technologies. This paper addresses data movement via an in-memory-computing accelerator that employs charged-domain mixed-signal operation for enhancing compute SNR and, thus, scalability. The architecture supports analog/binary input activation (IA)/weight first layer (FL) and binary/binary IA/weight hidden layers (HLs), with batch normalization and input–output (IO) (buffering) circuitry to enable cascading, if desired, for realizing different DNN layers. The architecture is arranged as $8\times 8=64$ in-memory-computing neuron tiles, supporting up to 512, $3\times 3\times 512$ -input HL neurons and 64, $3\times 3\times 3$ -input FL neurons, configurable via tile-level clock gating. In-memory computing is achieved using an 8T bit cell with overlaying metal-oxide-metal (MOM) capacitor, yielding a structure having $1.8\times $ the area of a standard 6T bit cell. Implemented in 65-nm CMOS, the design achieves HLs/FL energy efficiency of 866/1.25 TOPS/W and throughput of 18876/43.2 GOPS (1498/3.43 GOPS/mm2), when implementing convolution layers; and 658/0.95 TOPS/W, 9438/10.47 GOPS (749/0.83 GOPS/mm2), when implementing convolution followed by batch normalization layers. Several large-scale neural networks are demonstrated, showing performance on standard benchmarks (MNIST, CIFAR-10, and SVHN) equivalent to ideal digital computing.

A 64-Tile 2.4-Mb In-Memory-Computing CNN Accelerator Employing Charge-Domain Compute

A crossbar array of magnetoresistive memory devices for in-memory computing

In-memory computing (IMC) addresses the cost of accessing data from memory in a manner that introduces a tradeoff between energy/throughput and computation signal-to-noise ratio (SNR). However, low SNR posed a primary restriction to integrating IMC in larger, heterogeneous architectures required for practical workloads due to the challenges with creating robust abstractions necessary for the hardware and software stack. This work exploits recent progress in high-SNR IMC to achieve a programmable heterogeneous microprocessor architecture implemented in 65-nm CMOS and corresponding interfaces to the software that enables mapping of application workloads. The architecture consists of a 590-Kb IMC accelerator, configurable digital near-memory-computing (NMC) accelerator, RISC-V CPU, and other peripherals. To enable programmability, microarchitectural design of the IMC accelerator provides the integration in the standard processor memory space, area- and energy-efficient analog-to-digital conversion for interfacing to NMC, bit-scalable computation (1–8 b), and input-vector sparsity-proportional energy consumption. The IMC accelerator demonstrates excellent matching between computed outputs and idealized software-modeled outputs, at 1b TOPS/W of 192|400 and 1b-TOPS/mm2 of 0.60|0.24 for MAC hardware, at $V_{DD}$ of 1.2|0.85 V, both of which scale directly with the bit precision of the input vector and matrix elements. Software libraries developed for application mapping are used to demonstrate CIFAR-10 image classification with a ten-layer CNN, achieving accuracy, throughput, and energy of 89.3%|92.4%, 176|23 images/s, and $5.31\mid 105.2~\mu \text{J}$ /image, for 1|4 b quantization levels.

A Programmable Heterogeneous Microprocessor Based on Bit-Scalable In-Memory Computing

This paper presents a scalable neural-network (NN) inference accelerator in 16nm, based on an array of programmable cores employing mixed-signal In-Memory Computing (IMC), digital Near-Memory Computing (NMC), and localized buffering/control. IMC achieves high energy efficiency and throughput for matrix-vector multiplications (MVMs), which dominate NNs; but, scalability poses numerous challenges, both technologically, going to advanced nodes to maintain gains over digital architectures, and architecturally, for full execution of diverse NNs. Recent demonstrations have explored integrating IMC in programmable processors [1, 2], but have not achieved IMC efficiency and throughput for full executions. The central challenge is drastically different physical design points and associated tradeoffs incurred by IMC compared to digital engines. Namely, IMC substantially increases compute energy efficiency and HW density/parallelism, but retains the overheads of HW virtualization (state and data swapping/buffering/communication across spatial/temporal computation mappings). The demonstrated architecture is co-designed with SW-mapping algorithms (encapsulated in a custom graph compiler), to provide efficiency across a broad range of mapping strategies, to overcome these overheads.

A Programmable Neural-Network Inference Accelerator Based on Scalable In-Memory Computing

This paper presents an in-memory computing (IMC) macro in 28nm for fully row/column-parallel matrix-vector multiplication (MVM), exploiting precise capacitor-based analog computation to extend from binary input-vector elements to 5-b input-vector elements, for 16x increase in energy efficiency and 5x increase in throughput. The 1152(row)x256(col.) macro employs multi-level input drivers based on a digital-switch DAC implementation, which preserve compute accuracy well beyond the 8-b resolution of the output ADCs, and whose area is halved via a dynamic-range doubling (DRD) technique. The macro achieves the highest reported IMC energy efficiency of 5796 TOPS/W and compute density of 12 TOPS/mm2 (both normalized to 1-b ops). CIFAR-10 image classification is demonstrated with accuracy of 91%, equal to the level of ideal SW implementation.

Fully Row/Column-Parallel In-memory Computing SRAM Macro employing Capacitor-based Mixed-signal Computation with 5-b Inputs

In-memory computing is an emerging approach for overcoming memory-accessing bottlenecks, by eliminating the costs of explicitly moving data from point of storage to point of computation outside the array However, computation increases the dynamic range of signals, such that performing it via the existing structure of dense memory substantially squeezes the signal-to-noise ratio (SNR) In this brief, we explore how computations can be scaled up, to jointly optimize energy/latency/bandwidth gains with SNR requirements We employ algorithmic techniques to decompose computations so that they can be mapped to multiple parallel memory banks operating at chosen optimal points Specifically focusing on in-memory classification, we consider a custom IC in 130-nm CMOS IC and demonstrate an algorithm combining error-adaptive classifier boosting and multi-armed bandits, to enable segmentation of a feature vector into multiple subsets The measured performance of 10-way MNIST digit classification, using images downsampled to $16{\times }16$ pixels (mapped across four separate banks), is 91%, close to that simulated using full unsegmented feature vectors The energy per classification is 8797 pJ, $143{\times }$ lower than that of a system based on separated memory and digital accelerator

Yinqi Tang

Papers

In-Memory Computing: Advances and prospects

A Programmable Heterogeneous Microprocessor Based on Bit-Scalable In-Memory Computing

A Programmable Neural-Network Inference Accelerator Based on Scalable In-Memory Computing

Fully Row/Column-Parallel In-memory Computing SRAM Macro employing Capacitor-based Mixed-signal Computation with 5-b Inputs

Scaling Up In-Memory-Computing Classifiers via Boosted Feature Subsets in Banked Architectures