CLChuan-Tung Lin

Chuan-Tung Lin

*This is an algorithmically generated author record based on publicly available information.

Current Organizations / Institutes

The organizations/institutes from the scholar's latest article


Chuan-Tung Lin has a total of 11 co-authors. Here are the top co-authors by number of shared publications:


Latest Publications

  • STAR-SRAM: 16-bit Floating-Point SRAM-Based Digital Computing-in-Memory Macro in a 28 nm

    by Chuan-Tung Lin, Jonghyun Oh, Mingoo Seok

    A digital computing-in-memory (DCIM) macro emerges as a promising building block in a deep neural network (DNN) accelerator. To better support DNN workloads, circuit designers aim to improve three main metrics for macros: energy efficiency, compute density, and weight density. Improvements in those metrics directly translate into reduced energy consumption, shorter inference latency, and a smaller silicon footprint at the accelerator level. This article introduces STAR-SRAM, a DCIM macro designed for 16-bit brain floating-point (BF16) arithmetic. The macro is designed using a top-down approach. We started the design process with model-based parameter optimization, in which we determined the optimal design parameters that maximize the figure of merit (FoM) of a DCIM macro. Then, we developed several circuit techniques that further improve the FoM, including pre-multiply symmetric alignment, approximate multiplier circuits, and sparsity-aware wordline, clock, and input gating techniques. Fabricated in 28-nm CMOS, STAR-SRAM achieves the state-of-the-art FoM with up to 43.06 TFLOPS/W, 1.89 TFLOPS/mm2, and 400 Kb/mm2.


  • D6CIM: 60.4-TOPS/W All-Digital 6T-SRAM-Based Compute-in-Memory Macro Supporting 1-to-8 b Fixed-Point Arithmetic in a 28-nm CMOS

    by Jonghyun Oh, Chuan-Tung Lin, Mingoo Seok

    This article presents an all-digital 6T-SRAM-based compute-in-memory (D6CIM) macro that supports 1-to-8-bit fixed-point vector matrix multiplication. The D6CIM is fully digital, employing only digital standard cells and 6T SRAM bitcells. The D6CIM adopts a time-sharing architecture with an optimal degree of time-sharing to maximize the product of compute density and weight density. It also features several circuit techniques, namely non-precharge single-ended access, hybrid compressor adder-tree circuits, and bidirectional shift-accumulation. We prototype the D6CIM macro in a 28-nm CMOS technology. The maximum operating frequency is ~360 MHz in a 1.1-V supply. The energy efficiency is measured at up to 60.4 TOPS/W. It achieves a compute density of 1.46 TOPS/mm2 and a weight density of 1005 kb/mm2. Compared to prior state-of-the-art CIM macros, the D6CIM achieves 1.3-4.8X improvement in the product of energy efficiency, compute density, and weight density.


  • AACIM: A 2785-TOPS/W, 161-TOP/mm 2 , <1.17%-RMSE, Analog-In Analog-Out Computing-In-Memory Macro in 28 nm

    by Rentao Wan, Chuan-Tung Lin, Bo Zhang, Yannis Tsividis, Mingoo Seok

    This article presents an analog-in analog-out CIM macro (AACIM) for use in analog deep neural network (DNN) processors. Our macro receives analog inputs, performs a 64-by-32 vector–matrix multiplication (VMM) with a current-discharging computation mechanism, and produces analog outputs. It stores a 4-bit weight as an analog voltage in the custom 3T1C capacitor-based memory cell. The macro features a current-based self-calibration writing (CSCW) scheme to program weight-proportional currents into cells with a feedback loop, greatly reducing the writing error across process, voltage, and temperature (PVT) variations. In addition, the macro uses a discharging current stabilization (DCS) scheme to mitigate current degradation due to short-channel effects, improving the computation accuracy. Besides, it combines several techniques to mitigate coupling noise and leakages, extending the cell retention time to approximately $300~\mu $ s. Fabricated in a 28-nm process, the AACIM macro demonstrates an energy efficiency of 2785 TOPS/W and a compute density of 160.9 TOPS/mm2, with a weight density of 1.67 Mb/mm2 for 4-bit operations at 0.9 V. In addition, it maintains robust performance with an RMSE of less than 1.17% across temperature variations from $- 20~^{\circ }$ C to $70~^{\circ }$ C and ±10% supply voltage variations. We also confirm the functionality at a cryogenic temperature with a significantly improved retention time.


  • iMCU: A 28-nm Digital In-Memory Computing-Based Microcontroller Unit for TinyML

    by Chuan-Tung Lin, Paul Xuanyuanliang Huang, Jonghyun Oh, Dewei Wang, 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 $^2$ .


  • DIMCA: An Area-Efficient Digital In-Memory Computing Macro Featuring Approximate Arithmetic Hardware in 28 nm

    by Chuan-Tung Lin, Dewei Wang, Bo Zhang, Gregory K. Chen, Phil C. Knag, Ram Kumar Krishnamurthy, 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.