Dewei Wang

Dewei Wang

Auto-generated profile, based on publicly available information

Dewei Wang

Department of Electrical Engineering, Columbia University, New York, NY, USA

Research Focus & Metrics

Academic Career Length

13Years (2014-Present)

(Calculated once Scilit publication data is synced: current year − year of first publication + 1)

Representative Publications

View All Publications
28 September 2023 in IEEE Journal of Solid-State Circuits
Verified
Journal Article
Closed access
CLChuan-Tung Lin
Chuan-Tung Lin
DWDewei Wang
Dewei Wang
BZBo Zhang
Bo Zhang
GCGregory K. Chen
Gregory K. Chen
PKPhil C. Knag
Phil C. Knag
MSMingoo Seok
Mingoo Seok
Recent SRAM-based in-memory computing (IMC) hardware demonstrates high energy efficiency and throughput for matrix–vector multiplication (MVM), the dominant kernel for deep neural networks (DNNs). Earlier IMC macros have employed analog-mixed-signal (AMS) arithmetic hardware. However, those so-called AIMCs suffer from process, voltage, and temperature (PVT) variations. Digital IMC (DIMC) macros, on the other hand, exhibit better robustness against PVT variations, but they tend to require more silicon area. This article proposes novel DIMC hardware featuring approximate arithmetic (DIMCA) to improve area efficiency without hurting compute density (CD). We also propose an approximation-aware training model and a customized number format to compensate for the accuracy degradation caused by the approximation hardware. We prototyped the test chip in 28-nm CMOS. It contains two versions: the DIMCA with single-approximate hardware (DIMCA1) and DIMCA with double-approximate hardware (DIMCA2). The measurement results show that DIMCA1 supports a 4 b-activation and 1 b-weight (4 b/1 b) CNN model, achieving 327 kb/mm2, 458–990 TOPS/W (normalized to 1 b/1 b), 8.27–392 TOPS/mm2 (normalized to 1 b/1 b), and 90.41% accuracy for CIFAR-10. DIMCA2 supports a 1 b/1 b CNN model, achieving 485 kb/mm2, 932–2219 TOPS/W, 14.4–607 TOPS/mm2, and 86.96% accuracy for CIFAR-10.

14 February 2024 in IEEE Journal of Solid-State Circuits
Verified
Journal Article
Closed access
CLChuan-Tung Lin
Chuan-Tung Lin
PHPaul Xuanyuanliang Huang
Paul Xuanyuanliang Huang
JOJonghyun Oh
Jonghyun Oh
DWDewei Wang
Dewei Wang
MSMingoo Seok
Mingoo Seok
Tiny machine learning (TinyML) envisions executing a deep neural network (DNN)-based inference on an edge device for improving battery life, latency, security, and privacy. Toward this vision, recent microcontroller units (MCUs) integrate in-memory computing (IMC) hardware to leverage its high energy efficiency and throughput in vector–matrix multiplication (VMM). However, those existing works require large IMC hardware, severely increasing the area overhead. In addition, most existing works use analog–mixed-signal (AMS) IMC hardware, exhibiting limited robustness over process, voltage, and temperature (PVT) variations. Finally, none can support a practical software development framework such as TensorFlow Lite for Microcontrollers (TFLite-micro). Due to these limitations, those MCUs did not present the performance for the standard benchmark MLPerf-Tiny, which makes it difficult to evaluate them against the state-of-the-art neural (not necessarily IMC-based) MCUs. In this article, we design a new IMC-based MCU, titled iMCU, for TinyML to address those challenges. In the design process, we: 1) define the optimal set of acceleration targets and 2) devise an area-efficient computation flow that requires the least amount of IMC hardware yet still provides a significant acceleration. In addition, we develop: 1) state-of-the-art digital IMC macros and 2) create the accelerator based on the macros, which can support the proposed computation flow in a fully pipelined manner. Combining those innovations, we prototyped the iMCU in a 28-nm CMOS. The measurement results show that the iMCU significantly outperforms the prior IMC-based MCUs in compute density, energy efficiency, and SRAM density (total SRAM size/total SRAM area). It also achieves a compact footprint of 2.73 mm $^2$ .

17 October 2024 in IEEE Journal of Solid-State Circuits
Verified
Journal Article
Closed access
DWDewei Wang
Dewei Wang
SKSung Justin Kim
Sung Justin Kim
MYMinhao Yang
Minhao Yang
ALAurel A. Lazar
Aurel A. Lazar
MSMingoo Seok
Mingoo Seok
This article presents keyword spotting (KWS) hardware that uses analog signal processing (ASP) and division-based energy normalization (DN) to perform ultralow power KWS robustly across unknown background noise conditions. Most existing KWS systems employed noise-dependent training, limiting their accuracy under different noise conditions. Some other works used noise-independent training but required a larger classifier to learn more features under different noise cases. In this work, we prototyped the KWS hardware using two chips: a normalized acoustic feature extractor (NAFE) chip in a 65-nm CMOS and a spiking neural network (SNN) classifier chip in the same technology. The proposed hardware can perform an end-to-end KWS task while achieving robustness across various signal-to-noise ratios (SNRs) and different noise types. In addition, it can mitigate the effects of the process and temperature variations of the ASP circuits, further improving accuracy. The proposed hardware is also fully event-driven, minimizing power consumption when the input activity is low. The end-to-end KWS system consumes 220 to 570 nW across 0%–100% input activities and achieves 96.5%–88.0% accuracy across 20 to $-$ 5 dB SNRs on the HeySnips dataset.

11 September 2023
Verified
Conference Paper
Closed access
DWDewei Wang
Dewei Wang
JOJonghyun Oh
Jonghyun Oh
GCGregory K. Chen
Gregory K. Chen
PKPhil Knag
Phil Knag
RKRam K. Krishnamurthy
Ram K. Krishnamurthy
MSMingoo Seok
Mingoo Seok
We present an ultra-low-power automatic speech recognition (ASR) accelerator in 28nm. We have developed a bio-inspired neuron model with per-neuron trainable decay rates and thresholds. The ASR recurrent neural network, designed with the neuron model, requires $\sim$4X less weights than LSTM to achieve the same accuracy. We also created a digital 6T-SRAM-based compute-in-memory (DCIM) macro. The ASR accelerator leverages the DCIM’s high computing throughput to reduce the required clock frequency and supply voltage, thereby reducing the power consumption to $32\mu \mathrm{W}$ at the competitive inference accuracy on the TIMIT ASR benchmark.

1 July 2014
Verified
Conference Paper
Closed access
DWDewei Wang
Dewei Wang
JXJianping Xiong
Jianping Xiong
>Inverted echo sounder (IES) is a specific instrument for undersea exploration. It can be used to measure the transmission time of the acoustic wave...