Taegeun Yoo’s research while affiliated with Samsung and other places

In this work, we present a novel 8T static random access memory (SRAM)-based compute-in-memory (CIM) macro for processing neural networks with high energy efficiency. The proposed 8T bitcell is free from disturb issues thanks to the decoupled read channels by adding two extra transistors to the standard 6T bitcell. A 128 $\times$ 128 8T SRAM array offers massively parallel binary multiply and accumulate (MAC) operations with 64 $\times$ binary inputs (0/1) and 64 $\times$ 128 binary weights (+1/–1). After parallel MAC operations, 128 column-based neurons generate 128 $\times$ 1–5 bit outputs in parallel. The proposed column-based neuron comprises 64 $\times$ bitcells for dot-product, 32 $\times$ bitcells for analog-to-digital converter (ADC), and 32 $\times$ bitcells for offset calibration. The column ADC with 32 $\times$ replica SRAM bitcells converts the analog MAC results (i.e., a differential read bitline (RBL/RBLb) voltage) to the 1–5 bit output code by sweeping their reference levels in 1–31 cycles (i.e., $2^{N}$ –1 cycles for N -bit ADC). The measured linearity results [differential nonlinearity (DNL) and integral nonlinearity (INL)] are +0.314/–0.256 least significant bit (LSB) and + 0.27/–0.116 LSB, respectively, after offset calibration. The simulated image classification results are 96.37% for Mixed National Institute of Standards and Technology database (MNIST) using a multi-layer perceptron (MLP) with two hidden layers, 87.1%/82.66% for CIFAR-10 using VGG-like/ResNet-18 convolutional neural networks (CNNs), demonstrating slight accuracy degradations (0.67%–1.34%) compared with the software baseline. A test chip with a 16K 8T SRAM bitcell array is fabricated using a 65-nm process. The measured energy efficiency is 490–15.8 TOPS/W for 1–5 bit ADC resolution using 0.45-/0.8-V core supply.

This work proposes a 2-dimensional programable SRAM-based PUF. The selection of challenge groups, orders, and sequence lengths dominates the responses with challenge-response pairs (CRPs) by order of rows $^{\mathrm {(sequence~\textrm {}length- 1)}} \times$ columns $^{\mathrm {(sequence~\textrm {}length - 1)}}$ . The PUF bit cell has split word-lines with vertical and horizontal connections, the bit-lines are placed orthogonally to generate one-bit data with four cells, the entropy source is enriched to 24 transistors. The proposed PUF supports multiple data maps from a single chip. A test chip was fabricated in 65 nm CMOS technology. Under 0.8V and 20 °C (nominal point), the bit error rate reaches 3%. In a single chip, the hamming distance achieves 42.49% within the same group and different orders of challenges, and 47.32% within the different groups of challenges (when the sequence length is 5). The measured inter-hamming distance between chips is improved to 49.47%.

This article (Colonnade) presents a fully digital bit-serial compute-in-memory (CIM) macro. The digital CIM macro is designed for processing neural networks with reconfigurable 1–16 bit input and weight precisions based on bit-serial computing architecture and a novel all-digital bitcell structure. A column of bitcells forms a column MAC and used for computing a multiply-and-accumulate (MAC) operation. The column MACs placed in a row work as a single neuron and computes a dot-product, which is an essential building block of neural network accelerators. Several key features differentiate the proposed Colonnade architecture from the existing analog and digital implementations. First, its full-digital circuit implementation is free from process variation, noise susceptibility, and data-conversion overhead that are prevalent in prior analog CIM macros. A bitwise MAC operation in a bitcell is performed in the digital domain using a custom-designed XNOR gate and a full-adder. Second, the proposed CIM macro is fully reconfigurable in both weight and input precision from 1 to 16 bit. So far, most of the analog macros were used for processing quantized neural networks with very low input/weight precisions, mainly due to a memory density issue. Recent digital accelerators have implemented reconfigurable precisions, but they are inferior in energy efficiency due to significant off-chip memory access. We present a regular digital bitcell array that is readily reconfigured to a 1–16 bit weight-stationary bit-serial CIM macro. The macro computes parallel dot-product operations between the weights stored in memory and inputs that are serialized from LSB to MSB. Finally, the bit-serial computing scheme significantly reduces the area overhead while sacrificing latency due to bit-by-bit operation cycles. Based on the benefits of digital CIM, reconfigurability, and bit-serial computing architecture, the Colonnade can achieve both high performance and energy efficiency (i.e., both benefits of prior analog and digital accelerators) for processing neural networks. A test-chip with $128 \times 128$ SRAM-based bitcells for digital bit-serial computing is implemented using 65-nm technology and tested with 1–16 bit weight/input precisions. The measured energy efficiency is 117.3 TOPS/W at 1 bit and 2.06 TOPS/W at 16 bit.

A novel 4T2C ternary embedded DRAM (eDRAM) cell is proposed for computing a vector-matrix multiplication in the memory array. The proposed eDRAM-based compute-in-memory (CIM) architecture addresses a well-known Von Neumann bottle-neck in the traditional computer architecture and improves both latency and energy in processing neural networks. The proposed ternary eDRAM cell takes a smaller area than prior SRAM-based bitcells using 6–12 transistors. Nevertheless, the compact eDRAM cell stores a ternary state (−1, 0, or +1), while the SRAM bitcells can only store a binary state. We also present a method to mitigate the compute accuracy degradation issue due to device mismatches and variations. Besides, we extend the eDRAM cell retention time to $200~\mu \text{s}$ by adding a custom metal capacitor at the storage node. With the improved retention time, the overall energy consumption of eDRAM macro, including a regular refresh operation, is lower than most of prior SRAM-based CIM macros. A $128\times 128$ ternary eDRAM macro computes a vector-matrix multiplication between a vector with 64 binary inputs and a matrix with $64\times 128$ ternary weights. Hence, 128 outputs are generated in parallel. Note that both weight and input bit-precisions are programmable for supporting a wide range of edge computing applications with different performance requirements. The bit-precisions are readily tunable by assigning a variable number of eDRAM cells per weight or adding multiple pulses to input. An embedded column ADC based on replica cells sweeps the reference level for $2^{\mathrm {N}}-1$ cycles and converts the analog accumulated bitline voltage to a 1-5bit digital output. A critical bitline accumulate operation is simulated (Monte-Carlo, 3K runs). It shows the standard deviation of 2.84% that could degrade the classification accuracy of the MNIST dataset by 0.6% and the CIFAR-10 dataset by 1.3% versus a baseline with no variation. The simulated energy is 1.81fJ/operation, and the energy efficiency is 552.5-17.8TOPS/W (for 1-5bit ADC) at 200MHz using 65nm technology.

Gesture recognition has increasingly become one of the most popular human–machine interaction techniques for smart devices. Existing gesture recognition systems suffer from either excessive power consumption or large size, limiting their applications for ultralow-power wearable devices. This article presents an accurate area-efficient and low-power real-time gesture recognition system for smart wearable devices. The proposed work utilizes an accurate peak-based gesture classification engine with less memory and a low-resolution and low-power on-chip image sensor for achieving high area efficiency and low power. In addition, the feature extraction architecture removes fixed-pattern noises from the low-power on-chip image sensor for accuracy improvement and employs parallelism for recognition speed enhancement. Thus, the proposed system accomplishes accurate real-time gesture recognition for eight motion hand gestures with an average recognition accuracy of 90.6% and latency of 4.228 ms. Measurement results of a test chip fabricated in 65-nm CMOS demonstrate that the proposed system consumes $137.0~\mu \text{W}$ at 30 frames/s while occupying only 1.78 mm ² , which achieves the lowest power and smallest area among the recently reported gesture recognition systems.

In this paper, a low power asynchronous successive approximation register (SAR) analog-to-digital converter (ADC) involving the process, voltage, and temperature (PVT) compensation is presented. A proposed adaptive conversion time detection-and-control technique enhances the power efficiency, covering wide PVT variations. The proposed detection-and-control technique senses PVT variation in an aspect of conversion time, and adaptively controls the operation speed and power consumption. For PVT compensation, the proposed architecture includes the local supply/ground voltage. The local supply/ground voltage makes high |VGS| for transistors in the comparator and capacitive digital-to-analog converter switches, resulting in enhanced operation speed. However, when PVT condition changes to be favorable for the conversion speed, the |VGS| decreases for low power consumption. 30 chips were measured to verify the proposed ADC. Having the proposed architecture tested with 10 kHz input frequency, SNDR remained higher than 60 dB at unfavorable conditions such as -9 % supply voltage variation, or -20 ⁰C temperature variation. On the other hand, at favorable conditions such as +9 % supply voltage variation, or 80 ⁰C temperature variation, the power consumption of SAR ADC decreased without performance degradation.

In Space applications, the scaling of transistors has made integrated circuits (ICs) more susceptible to soft errors, caused by radiation strikes. When a soft error causes a bit flip in a memory device, this event is referred to as a Single Event Upset (SEU). Since SEU errors degrade system performance and eventually lead to system failure, the design of radiation-resilient memory is substantial. This paper presents a radiation resilient SRAM with a self-refresh scheme for lowering the number of errors in each row below a threshold number. The proposed self-refresh operation reads out the stored data and performs single error correction using a simple algorithm during its hold/idle mode. A 4KB SRAM test chip in 65nm CMOS technology demonstrates a significant reduction in errors with the self-refresh operation. When the SRAM test chip was exposed to accelerated proton radiation with an energy level of 39.38 MeV, the self-refresh scheme reduces the number of uncorrectable errors by $25\times$ and $8\times$ lesser for the fluence of $9.82\times 10^{11}$ particles/cm ² and $49.1\times 10^{11}$ particles/cm ² , respectively.

This article presents an 8T static random access memory (SRAM) macro with vertical read wordline (RWL) and selective dual split power (SDSP) lines techniques. The proposed vertical RWL reduces dynamic energy consumption during read operation by charging and discharging only selected read bitlines (RBLs). The data-aware SDSP technique combined with vertical write bitlines enhances both the write margin (WM) and the static noise margin (SNM). A 16-kb SRAM test chip fabricated in 65-nm CMOS technology demonstrates the minimum energy consumption of 0.506 pJ at 0.4 V and the minimum operating voltage of 0.26 V.