Search

Convolutional neural networks (CNNs) play a key role in deep learning applications. However, the large storage overheads and the substantial computation cost of CNNs are problematic in hardware accelerators. Computing-in-memory (CIM) architecture has demonstrated great potential to effectively compute large-scale matrix-vector multiplication. However, the intensive multiply and accumulation (MAC) operations executed at the crossbar array and the limited capacity of CIM macros remain bottlenecks for further improvement of energy efficiency and throughput. To reduce computation costs, network pruning and quantization are two widely studied compression methods to shrink the model size. However, most of the model compression algorithms can only be implemented in digital-based CNN accelerators. For implementation in a static random access memory (SRAM) CIM-based accelerator, the model compression algorithm must consider the hardware limitations of CIM macros, such as the number of word lines and bit lines that can be turned on at the same time, as well as how to map the weight to the SRAM CIM macro. In this study, a software and hardware co-design approach is proposed to design an SRAM CIM-based CNN accelerator and an SRAM CIM-aware model compression algorithm. To lessen the high-precision MAC required by batch normalization (BN), a quantization algorithm that can fuse BN into the weights is proposed. Furthermore, to reduce the number of network parameters, a sparsity algorithm that considers a CIM architecture is proposed. Last, MARS, a CIM-based CNN accelerator that can utilize multiple SRAM CIM macros as processing units and support a sparsity neural network, is proposed., IEEE Transactions on Computer-Aided Design of Integrated Circuits and Systems 2021

Computing-in-memory (CIM) based on SRAM is a promising approach to achieving energy-efficient multiply-and-accumulate (MAC) operations in artificial intelligence (AI) edge devices; however, existing SRAM-CIM chips support only DNN inference. The flow of training data requires that CIM arrays perform convolutional computation using transposed weight matrices. This article presents a two-way transpose (TWT) multiply cell with high resistance to process variation and a novel read scheme that uses input-aware zone prediction of maximum partial MAC values to enhance the signal margin for robust readout. A 28-nm 64-kb TWT CIM macro fabricated using foundry-provided compact 6T-SRAM cells achieved ${T_{AC}}$ of 3.8-21 ns and energy efficiency of 7-61.1 TOPS/W in performing MAC operations using 2-8-b inputs, 4-8-b weights, and 10-20-b outputs.

Computing-in-memory (CIM) architectures have gained importance in achieving high-throughput energy-efficient artificial intelligence (AI) systems. Resistive RAM (RRAM) is a promising candidate for CIM architectures due to a multiply-and-accumulate (MAC)-friendly structure, high bit density, compatibility with a CMOS process, and nonvolatility. Notwithstanding the advancement of RRAM technology, the reliability of an RRAM array hinders the spread of RRAM applications such that a circuit-technology joint approach is necessary to attain reliable RRAM-based CIM architectures. This article presents a 64-kb hybrid CIM/digital RRAM macro supporting: 1) active-feedback-based voltage-sensing read (RD) to enable 1-8-b programmable vector-matrix multiplication under a low-resistance ratio of the high-resistance state to the low-resistance state in an RRAM array; 2) iterative write with verification to secure a tight resistance distribution; and 3) online RD-disturb detection in the background during CIM. The test chip fabricated in a 40-nm CMOS and RRAM process achieves a peak energy efficiency of 56.67 TOPS/W while demonstrating the eight-bitline hybrid CIM/digital MAC operation with 1-8-b inputs and weights and 20-b outputs without quantization.

Learning from a few examples (one/few-shot learning) on the fly is a key challenge for on-device machine intelligence. We present the first chip-level demonstration of one-shot learning with Stanford Associative memory for Programmable, Integrated Edge iNtelligence via life-long learning and Search (SAPIENS), a resistive random access memory (RRAM)-based non-volatile associative memory (AM) chip that serves as the backend for memory-augmented neural networks (MANNs). The 64-kb fully integrated RRAM-CMOS AM chip performs long-term feature embedding and retrieval, demonstrated on a 32-way one-shot learning task on the Omniglot dataset. Using only one example per class for 32 unseen classes during on-chip learning, SAPIENS achieves 79% measured inference accuracy on Omniglot, comparable to edge software model accuracy using five-level quantization (82%). It achieves an energy efficiency of 118 GOPS/W at 200 MHz for in-memory L1 distance computation and prediction. Multi-bank measurements on the same chip show that increasing the capacity from three banks (24 kb) to eight banks (64 kb) improves the chip accuracy from 73.5% to 79%, while minimizing the accuracy excursion due to bank-to-bank variability.

In this work, we present a four-transistor-two-resistor (4T2R) ternary content addressable memory (TCAM) bit cell based on the resistive memory (RRAM), comprising the conventional two-transistor-two-resistor (2T2R) cell with two additional comparison transistors. It can effectively amplify the match-line signal ratio (ML-ratio), lower the leakage current of the match cell ( ${I}_{{\mathrm {MATCH}}}$ ), and suppress the read disturbance. The proposed concept is silicon verified using the 180 nm CMOS technology with transition-metal-oxide (TMO) RRAM integrated at the back-end-of-line (BEOL). It achieves a considerable ML-ratio of 1860 and a low ${I}_{{\mathrm {MATCH}}}$ of 11.15 nA on average. In a typical search operation, it shows a negligible ML drop at the match case and a large ML swing range at the mismatch case. The SPICE simulation results further show it can support a long word-length (WDL) of 256 under a clock rate of 100 MHz for search operations, which demonstrates its promise for highly parallel nonvolatile TCAM.

This article presents a computing-in-memory (CIM) structure aimed at improving the energy efficiency of edge devices running multi-bit multiply-and-accumulate (MAC) operations. The proposed scheme includes a 6T SRAM-based CIM (SRAM-CIM) macro capable of: 1) weight-bitwise MAC (WbwMAC) operations to expand the sensing margin and improve the readout accuracy for high-precision MAC operations; 2) a compact 6T local computing cell to perform multiplication with suppressed sensitivity to process variation; 3) an algorithm-adaptive low MAC-aware readout scheme to improve energy efficiency; 4) a bitline header selection scheme to enlarge signal margin; and 5) a small-offset margin-enhanced sense amplifier for robust read operations against process variation. A fabricated 28-nm 64-kb SRAM-CIM macro achieved access times of 4.1–8.4 ns with energy efficiency of 11.5–68.4 TOPS/W, while performing MAC operations with 4- or 8-b input and weight precision.

This article presents a low-power, high-speed smart vision sensor for motion detection (MD) that realizes in-pixel frame-difference (FD) operation using a global shutter mechanism. A ping-pong pulse-width-modulation (PWM) pixel is proposed to achieve the consecutive event frame report with a balanced signal transfer function of successive FD operations. Three operating modes were implemented for varied application scenarios, such as image capture (IC) mode to capture a raw image, FD mode for MD, and saliency detection (SD) mode for low-resolution sub-block event counting. A 0.8 V 64 $\times $ 64 vision sensor prototype was fabricated and verified in TSMC 0.18- $\mu \text{m}$ standard CMOS technology. In IC mode, it consumed 71.2 $\mu \text{W}$ @360fps with an achieved iFoM of 48.3 $\text {pJ/pixel}\cdot \text {frame}$ . In FD mode, it consumed 74.4 $\mu \text{W}$ @510fps with full-resolution (64 $\times $ 64) event reporting and achieved iFoMs of 35.6 $\text {pJ/pixel}\cdot \text {frame}$ . In SD mode, it consumed 121.6 $\mu \text{W}$ @890fps with block-level (8 $\times $ 8) saliency reporting and achieved iFoMs of 2.1 $\text {nJ/block}\cdot \text {frame}$ .

The emerging edge intelligence requires low-cost energy-efficient neural network (NN) processors. Supporting various types of edge NN models leads to extra circuit overhead. Designing a unified NN processor with high energy/area efficiency is challenging. This work presents a frequency-domain-accelerated unified NN processor, named STICKER-T. It combines algorithm, architecture, and circuit-level optimization to achieve high energy/area efficiency. By utilizing the block-circulant NN (CirCNN) algorithm, this work supports frequency-domain acceleration and a unified workflow for convolutional, fully connected, and recurrent NN (CNN/FC/RNN). Three key innovations are proposed. First, a block-circulant-accelerated chip architecture is implemented to support unified CNN/FC/RNN workflow. Second, a multi-bit 8-128-point global-parallel local-bit-serial fast Fourier transform (FFT) module is designed for efficient high-throughput FFT/inverse FFT (IFFT) operation. Third, by utilizing a 6T hierarchical-bitline-switching transpose-SRAM (HBST-TRAM), 2-D data reuse is enabled in the proposed multi-bit frequency-domain multiply–accumulate (MAC) array. STICKER-T was fabricated in a 65-nm CMOS technology. It can operate at 0.54–1.15 V and 25–200 MHz with 13.3–339-mW power consumption. The peak energy efficiency achieves 140.3 TOPS/W. It shows 8.1 $\times $ area efficiency and 4.2 $\times $ energy efficiency at 4-bit precision compared with the state-of-the-art reconfigurable NN processor.

The compare operation is widely used in many applications, from fundamental sorting to primitive operations in the database and AI systems. We present SRAM-based 3-D-CAM circuit designs using a monolithic 3-D (M3D) integration process for realizing beyond-Boolean in-memory compare operation without any area overheads. We also fabricated a processing-in-memory (PiM) macro with the same 3-D-CAM circuit using M3D for performing massively parallel compare operations used in the database, machine learning, and scientific applications. We show various system designs with the 3-D-CAM supporting operations, such as data filtering, sorting, and sparse matrix–matrix multiplication (SpGEMM). Our systems exhibit up to $272\times $ , $200\times $ , and $226\times $ speedups and $151\times $ , $37\times $ , and $156\times $ energy savings compared to systems using near memory compute for the data filtering, sorting, and SpGEMM applications, respectively.

When applied to artificial intelligence edge devices, the conventionally von Neumann computing architecture imposes numerous challenges (e.g., improving the energy efficiency), due to the memory-wall bottleneck involving the frequent movement of data between the memory and the processing elements (PE). Computing-in-memory (CIM) is a promising candidate approach to breaking through this so-called memory wall bottleneck. SRAM cells provide unlimited endurance and compatibility with state-of-the-art logic processes. This paper outlines the background, trends, and challenges involved in the further development of SRAM-CIM macros. This paper also reviews recent silicon-verified SRAM-CIM macros designed for logic and multiplication-accumulation (MAC) operations.

Physically unclonable function (PUF) has been increasingly used as a promising primitive for hardware security with a wide range of applications in the Internet of Things (IoT). In recent years, novel PUF techniques based on resistive switching mechanism in various emerging nonvolatile memories have demonstrated superior performance on reliability and integration density. In this work, a resistive random access memory (RRAM)-based PUF chip with 8-kb capacity is developed. Two operation modes, namely differential mode and median mode, are embedded on chip. To implement these modes, a current sampling-based sense amplifier is designed to distinguish the current values of the PUF cells and the reference cell. In addition, a split-resistance scheme is proposed to enhance the PUF’s reliability significantly. The experiment results show that the differential PUF exhibits excellent performance with native bit error rate (N-BER) below 6 $\times $ 10−6 and inter-Hamming distance (inter-HD) of 49.99%. In the meanwhile, the reconfigurability of PUF challenge-response pairs (CRPs) is demonstrated with 49.77% and 47.29% reconfigure-Hamming distance (reconfigure-HD) in the median mode and the differential mode, respectively.

Resistive memory (RRAM) provides an ideal platform to develop embedded non-volatile computing-in-memory (nvCIM). However, it faces several critical challenges ranging from device non-idealities, large DC currents, and small signal margins. To address these issues, we propose voltage-division (VD) based computing approach and its circuit implementation in two-transistor-two-resistor (2T2R) RRAM cell arrays, which can realize energy-efficient, sign-aware, and robust deep neural network (DNN) processing. A readout technique, namely the input-dependent sensing control (IDSC) scheme, is also introduced for power saving. On this basis, a 400kb VD-based RRAM nvCIM is silicon verified. It achieves 2.54 times power reduction compared to that of the ones rely on conventional weighted-current summation (WCS) mechanism, a peak energy-efficiency of 42.6 TOPS/W and a minimum latency of 15.98 ns.

As the growing demand on artificial intelligence (AI) Internet-of-Things (IoT) devices, smart vision sensors with energy-efficient computing capability are required. This article presents a low-power and low-voltage dual mode 0.5-V computational CMOS image sensor (C2IS) with array-parallel computing capability for feature extraction using convolution. In the feature extraction mode, by applying the pulsewidth modulation (PWM) pixel and switch-current integration (SCI) circuit, the in-sensor eight-directional matrix-parallel multiply–accumulate (MAC) operation is realized. Furthermore, the analog-domain convolution-on-readout (COR) operation, the programmable $3\times3$ kernel with ±3-bit weights, and the tunable-resolution column-parallel analog-to-digital converter (ADC) (1–8 bit) are implemented to achieve the real-time feature extraction without using additional memory and sacrificing frame rate. In the image capturing mode, the sensor provides the linear-response 8-bit raw image data. The C2IS prototype has been fabricated in the TSMC 0.18- $\mu \text{m}$ standard process technology and verified to demonstrate the raw and feature images at 480 frames/s with a power consumption of 77/ $117~\mu \text{W}$ and the resultant FoM of 9.8/14.8 pJ/pixel/frame, respectively. The prototype sensor is used as a real-time edge feature detection frond-end camera and accompanied with a simplified convolutional neural network (CNN) architecture to demonstrate the hand gesture recognition. The prototype system achieves more than 95% validation accuracy.