18 March 2021 in Nature Nanotechnology
2 July 2019 in IEEE Electron Device Letters
Phase change memory (PCM) has been considered as one of the most promising emerging non-volatile memories for in-memory computing of neural networks. In this letter, we investigate the impact of resistance drift and its statistical variations on two widely-used artificial neural network (ANN) models, multi-layer perceptron (MLP) and convolutional neural network (CNN). We employ experimentally measured resistance drift characteristics into the ANN models to accurately model weight updates represented by PCM synaptic devices. Our results suggest that the resistance drift in PCM causes minor accuracy degradation (only ∼1%) for both MLP and CNN models. However, classification accuracy can be significantly reduced if the PCM drift characteristics exhibit high device-to-device and cycle-to-cycle variations in the drift coefficients.
27 September 2018 in IEEE Electron Device Letters
Phase change memory (PCM), one of the most mature emerging non-volatile memories, has gained considerable attention over the years for use as electronic synapses in biologically inspired neuromorphic systems. The resistance drift of PCM devices, nonetheless, has long been identified as one of the biggest challenges toward realizing many areas of applications. Although this drawback has been extensively studied for memory development and many methods were proposed to mitigate the drift effect, its impact, if any, on online learning has not been fully explored yet. In this letter, we investigate the impact of resistance drift and variations in resistance drift parameters during unsupervised online learning. We use the resistance drift characteristics measured from experiments and incorporate them into the spiking neural network (SNN) for MNIST handwritten digits classification. Our results show that resistance drift, considered as a non-ideality for PCM devices, can be exploited to boost accuracy for online learning of handwritten digits in the SNN.
20 May 2020 in IEEE Electron Device Letters
In recent years, Spin-Transfer-Torque Magnetic Random Access Memory (STT-MRAM) has been considered as one of the most promising non-volatile memory candidates for in-memory computing. However, system-level performance gains using STT-MRAM for in-memory computing at deeply scaled nodes have not been assessed with respect to more mature memory technologies. In this letter, we present perpendicular magnetic tunnel junction (pMTJ) STT-MRAM devices at 28nm and 7nm. We evaluate the system-level performance of convolutional neural network (CNN) inference with STT-MRAM arrays in comparison to Static Random Access Memory (SRAM). We benchmark STT-MRAM and SRAM in terms of area, leakage power, energy, and latency from 65nm to 7nm technology nodes. Our results show that STT-MRAM keeps providing ~5× smaller synaptic core area, ~20× less leakage power, and ~7× less energy than SRAM when both devices are scaled from 65nm to 7nm. With the emerging need for low power computation for a broad range of applications such as internet-of-things (IoT) and neural network (NN), STT-MRAM can offer energy-efficient and high-density in-memory computing.
13 December 2021 in IEEE Transactions on Electron Devices
Real-time spike sorting and processing are crucial for closed-loop brain–machine interfaces and neural prosthetics. Recent developments in high-density multielectrode arrays with hundreds of electrodes have enabled simultaneous recordings of spikes from a large number of neurons. However, the high channel count imposes stringent demands on real-time spike sorting hardware (HW) regarding data transmission bandwidth and computation complexity. Thus, it is necessary to develop a specialized real-time HW that can sort neural spikes on the fly with high throughputs while consuming minimal power. Here, we present a real-time, low latency spike sorting processor that utilizes high-density CuO x resistive crossbars to implement in-memory spike sorting in a massively parallel manner. We developed a fabrication process that is compatible with CMOS back end of line (BEOL) integration. We extensively characterized switching characteristics and statistical variations of the CuO x memory devices. In order to implement spike sorting with crossbar arrays, we developed a template matching-based spike sorting algorithm that can be directly mapped onto resistive random-access memory (RRAM) crossbars. By using synthetic and in vivo recordings of extracellular spikes, we experimentally demonstrated energy-efficient spike sorting with high accuracy. Our neuromorphic interface offers substantial improvements in area ( $\sim 1000\times $ less area), power ( $\sim 200\times $ less power), and latency ( $4.8~\mu \text{s}$ latency for sorting 100 channels) for real-time spike sorting compared to other HW implementations based on field-programmable gate arrays (FPGAs) and microcontrollers.