Xiaobo Sharon Hu

Jing Gong, Hassaan Saadat, Hasindu Gamaarachchi et al.

Edge training of Deep Neural Networks (DNNs) is a desirable goal for continuous learning; however, it is hindered by the enormous computational power required by training. Hardware approximate multipliers have shown their effectiveness for gaining resource-efficiency in DNN inference accelerators; however, training with approximate multipliers is largely unexplored. To build resource efficient accelerators with approximate multipliers supporting DNN training, a thorough evaluation of training convergence and accuracy for different DNN architectures and different approximate multipliers is needed. This paper presents ApproxTrain, an open-source framework that allows fast evaluation of DNN training and inference using simulated approximate multipliers. ApproxTrain is as user-friendly as TensorFlow (TF) and requires only a high-level description of a DNN architecture along with C/C++ functional models of the approximate multiplier. We improve the speed of the simulation at the multiplier level by using a novel LUT-based approximate floating-point (FP) multiplier simulator on GPU (AMSim). ApproxTrain leverages CUDA and efficiently integrates AMSim into the TensorFlow library, in order to overcome the absence of native hardware approximate multiplier in commercial GPUs. We use ApproxTrain to evaluate the convergence and accuracy of DNN training with approximate multipliers for small and large datasets (including ImageNet) using LeNets and ResNets architectures. The evaluations demonstrate similar convergence behavior and negligible change in test accuracy compared to FP32 and bfloat16 multipliers. Compared to CPU-based approximate multiplier simulations in training and inference, the GPU-accelerated ApproxTrain is more than 2500x faster. Based on highly optimized closed-source cuDNN/cuBLAS libraries with native hardware multipliers, the original TensorFlow is only 8x faster than ApproxTrain.

Zheyu Yan, Xiaobo Sharon Hu, Yiyu Shi

Computing-in-Memory (CiM) architectures based on emerging non-volatile memory (NVM) devices have demonstrated great potential for deep neural network (DNN) acceleration thanks to their high energy efficiency. However, NVM devices suffer from various non-idealities, especially device-to-device variations due to fabrication defects and cycle-to-cycle variations due to the stochastic behavior of devices. As such, the DNN weights actually mapped to NVM devices could deviate significantly from the expected values, leading to large performance degradation. To address this issue, most existing works focus on maximizing average performance under device variations. This objective would work well for general-purpose scenarios. But for safety-critical applications, the worst-case performance must also be considered. Unfortunately, this has been rarely explored in the literature. In this work, we formulate the problem of determining the worst-case performance of CiM DNN accelerators under the impact of device variations. We further propose a method to effectively find the specific combination of device variation in the high-dimensional space that leads to the worst-case performance. We find that even with very small device variations, the accuracy of a DNN can drop drastically, causing concerns when deploying CiM accelerators in safety-critical applications. Finally, we show that surprisingly none of the existing methods used to enhance average DNN performance in CiM accelerators are very effective when extended to enhance the worst-case performance, and further research down the road is needed to address this problem.

Zheyu Yan, Yifan Qin, Xiaobo Sharon Hu et al.

Deep Neural Networks (DNNs) have demonstrated impressive performance across a wide range of tasks. However, deploying DNNs on edge devices poses significant challenges due to stringent power and computational budgets. An effective solution to this issue is software-hardware (SW-HW) co-design, which allows for the tailored creation of DNN models and hardware architectures that optimally utilize available resources. However, SW-HW co-design traditionally suffers from slow optimization speeds because their optimizers do not make use of heuristic knowledge, also known as the ``cold start'' problem. In this study, we present a novel approach that leverages Large Language Models (LLMs) to address this issue. By utilizing the abundant knowledge of pre-trained LLMs in the co-design optimization process, we effectively bypass the cold start problem, substantially accelerating the design process. The proposed method achieves a significant speedup of 25x. This advancement paves the way for the rapid and efficient deployment of DNNs on edge devices.

Zheyu Yan, Yifan Qin, Wujie Wen et al.

Compute-in-Memory (CiM), built upon non-volatile memory (NVM) devices, is promising for accelerating deep neural networks (DNNs) owing to its in-situ data processing capability and superior energy efficiency. Unfortunately, the well-trained model parameters, after being mapped to NVM devices, can often exhibit large deviations from their intended values due to device variations, resulting in notable performance degradation in these CiM-based DNN accelerators. There exists a long list of solutions to address this issue. However, they mainly focus on improving the mean performance of CiM DNN accelerators. How to guarantee the worst-case performance under the impact of device variations, which is crucial for many safety-critical applications such as self-driving cars, has been far less explored. In this work, we propose to use the k-th percentile performance (KPP) to capture the realistic worst-case performance of DNN models executing on CiM accelerators. Through a formal analysis of the properties of KPP and the noise injection-based DNN training, we demonstrate that injecting a novel right-censored Gaussian noise, as opposed to the conventional Gaussian noise, significantly improves the KPP of DNNs. We further propose an automated method to determine the optimal hyperparameters for injecting this right-censored Gaussian noise during the training process. Our method achieves up to a 26% improvement in KPP compared to the state-of-the-art methods employed to enhance DNN robustness under the impact of device variations.

Xinzhao Li, Alptekin Vardar, Franz Müller et al.

Deploying Retrieval-Augmented Generation (RAG) on edge devices is in high demand, but is hindered by the latency of massive data movement and computation on traditional architectures. Compute-in-Memory (CiM) architectures address this bottleneck by performing vector search directly within their crossbar structure. However, CiM's adoption for RAG is limited by a fundamental ``representation gap,'' as high-precision, high-dimension embeddings are incompatible with CiM's low-precision, low-dimension array constraints. This gap is compounded by the diversity of CiM implementations (e.g., SRAM, ReRAM, FeFET), each with unique designs (e.g., 2-bit cells, 512x512 arrays). Consequently, RAG data must be naively reshaped to fit each target implementation. Current data shaping methods handle dimension and precision disjointly, which degrades data fidelity. This not only negates the advantages of CiM for RAG but also confuses hardware designers, making it unclear if a failure is due to the circuit design or the degraded input data. As a result, CiM adoption remains limited. In this paper, we introduce CQ-CiM, a unified, hardware-aware data shaping framework that jointly learns Compression and Quantization to produce CiM-compatible low-bit embeddings for diverse CiM designs. To the best of our knowledge, this is the first work to shape data for comprehensive CiM usage on RAG.

Yifan Qin, Jiahao Zheng, Zheyu Yan et al.

Compute-in-memory (CiM) architectures promise significant improvements in energy efficiency and throughput for deep neural network acceleration by alleviating the von Neumann bottleneck. However, their reliance on emerging non-volatile memory devices introduces device-level non-idealities-such as write variability, conductance drift, and stochastic noise-that fundamentally challenge reliability, predictability, and safety, especially in safety-critical applications. This talk examines the reliability limits of CiM-based neural accelerators and presents a series of techniques that bridge device physics, architecture, and learning algorithms to address these challenges. We first demonstrate that even small device variations can lead to disproportionately large accuracy degradation and catastrophic failures in safety-critical inference workloads, revealing a critical gap between average-case evaluations and worst-case behavior. Building on this insight, we introduce SWIM, a selective write-verify mechanism that strategically applies verification only where it is most impactful, significantly improving reliability while maintaining CiM's efficiency advantages. Finally, we explore a learning-centric solution that improves realistic worst-case performance by training neural networks with right-censored Gaussian noise, aligning training assumptions with hardware-induced variability and enabling robust deployment without excessive hardware overhead. Together, these works highlight the necessity of cross-layer co-design for CiM accelerators and provide a principled path toward dependable, efficient neural inference on emerging memory technologies-paving the way for their adoption in safety- and reliability-critical systems.

Yifan Qin, Zheyu Yan, Dailin Gan et al.

Compute-in-memory accelerators built upon non-volatile memory devices excel in energy efficiency and latency when performing deep neural network (DNN) inference, thanks to their in-situ data processing capability. However, the stochastic nature and intrinsic variations of non-volatile memory devices often result in performance degradation during DNN inference. Introducing these non-ideal device behaviors in DNN training enhances robustness, but drawbacks include limited accuracy improvement, reduced prediction confidence, and convergence issues. This arises from a mismatch between the deterministic training and non-deterministic device variations, as such training, though considering variations, relies solely on the model's final output. In this work, inspired by control theory, we propose Negative Feedback Training (NeFT), a novel concept supported by theoretical analysis, to more effectively capture the multi-scale noisy information throughout the network. We instantiate this concept with two specific instances, oriented variational forward (OVF) and intermediate representation snapshot (IRS). Based on device variation models extracted from measured data, extensive experiments show that our NeFT outperforms existing state-of-the-art methods with up to a 45.08% improvement in inference accuracy while reducing epistemic uncertainty, boosting output confidence, and improving convergence probability. These results underline the generality and practicality of our NeFT framework for increasing the robustness of DNNs against device variations. The source code for these two instances is available at https://github.com/YifanQin-ND/NeFT_CIM

Yifan Qin, Zheyu Yan, Zixuan Pan et al.

Compute-in-memory (CIM) accelerators using non-volatile memory (NVM) devices offer promising solutions for energy-efficient and low-latency Deep Neural Network (DNN) inference execution. However, practical deployment is often hindered by the challenge of dealing with the massive amount of model weight parameters impacted by the inherent device variations within non-volatile computing-in-memory (NVCIM) accelerators. This issue significantly offsets their advantages by increasing training overhead, the time and energy needed for mapping weights to device states, and diminishing inference accuracy. To mitigate these challenges, we propose the "Tiny Shared Block (TSB)" method, which integrates a small shared 1x1 convolution block into the DNN architecture. This block is designed to stabilize feature processing across the network, effectively reducing the impact of device variation. Extensive experimental results show that TSB achieves over 20x inference accuracy gap improvement, over 5x training speedup, and weights-to-device mapping cost reduction while requiring less than 0.4% of the original weights to be write-verified during programming, when compared with state-of-the-art baseline solutions. Our approach provides a practical and efficient solution for deploying robust DNN models on NVCIM accelerators, making it a valuable contribution to the field of energy-efficient AI hardware.

Zheyu Yan, Xiaobo Sharon Hu, Yiyu Shi

Architectures that incorporate Computing-in-Memory (CiM) using emerging non-volatile memory (NVM) devices have become strong contenders for deep neural network (DNN) acceleration due to their impressive energy efficiency. Yet, a significant challenge arises when using these emerging devices: they can show substantial variations during the weight-mapping process. This can severely impact DNN accuracy if not mitigated. A widely accepted remedy for imperfect weight mapping is the iterative write-verify approach, which involves verifying conductance values and adjusting devices if needed. In all existing publications, this procedure is applied to every individual device, resulting in a significant programming time overhead. In our research, we illustrate that only a small fraction of weights need this write-verify treatment for the corresponding devices and the DNN accuracy can be preserved, yielding a notable programming acceleration. Building on this, we introduce USWIM, a novel method based on the second derivative. It leverages a single iteration of forward and backpropagation to pinpoint the weights demanding write-verify. Through extensive tests on diverse DNN designs and datasets, USWIM manifests up to a 10x programming acceleration against the traditional exhaustive write-verify method, all while maintaining a similar accuracy level. Furthermore, compared to our earlier SWIM technique, USWIM excels, showing a 7x speedup when dealing with devices exhibiting non-uniform variations.

Zheyu Yan, Xiaobo Sharon Hu, Yiyu Shi

Emerging non-volatile memory (NVM)-based Computing-in-Memory (CiM) architectures show substantial promise in accelerating deep neural networks (DNNs) due to their exceptional energy efficiency. However, NVM devices are prone to device variations. Consequently, the actual DNN weights mapped to NVM devices can differ considerably from their targeted values, inducing significant performance degradation. Many existing solutions aim to optimize average performance amidst device variations, which is a suitable strategy for general-purpose conditions. However, the worst-case performance that is crucial for safety-critical applications is largely overlooked in current research. In this study, we define the problem of pinpointing the worst-case performance of CiM DNN accelerators affected by device variations. Additionally, we introduce a strategy to identify a specific pattern of the device value deviations in the complex, high-dimensional value deviation space, responsible for this worst-case outcome. Our findings reveal that even subtle device variations can precipitate a dramatic decline in DNN accuracy, posing risks for CiM-based platforms in supporting safety-critical applications. Notably, we observe that prevailing techniques to bolster average DNN performance in CiM accelerators fall short in enhancing worst-case scenarios. In light of this issue, we propose a novel worst-case-aware training technique named A-TRICE that efficiently combines adversarial training and noise-injection training with right-censored Gaussian noise to improve the DNN accuracy in the worst-case scenarios. Our experimental results demonstrate that A-TRICE improves the worst-case accuracy under device variations by up to 33%.

Zheyu Yan, Xiaobo Sharon Hu, Yiyu Shi

Computing-in-Memory architectures based on non-volatile emerging memories have demonstrated great potential for deep neural network (DNN) acceleration thanks to their high energy efficiency. However, these emerging devices can suffer from significant variations during the mapping process i.e., programming weights to the devices), and if left undealt with, can cause significant accuracy degradation. The non-ideality of weight mapping can be compensated by iterative programming with a write-verify scheme, i.e., reading the conductance and rewriting if necessary. In all existing works, such a practice is applied to every single weight of a DNN as it is being mapped, which requires extensive programming time. In this work, we show that it is only necessary to select a small portion of the weights for write-verify to maintain the DNN accuracy, thus achieving significant speedup. We further introduce a second derivative based technique SWIM, which only requires a single pass of forward and backpropagation, to efficiently select the weights that need write-verify. Experimental results on various DNN architectures for different datasets show that SWIM can achieve up to 10x programming speedup compared with conventional full-blown write-verify while attaining a comparable accuracy.

Dayane Reis, Haoran Geng, Michael Niemier et al.

This paper proposes IMCRYPTO, an in-memory computing (IMC) fabric for accelerating AES encryption and decryption. IMCRYPTO employs a unified structure to implement encryption and decryption in a single hardware architecture, with combined (Inv)SubBytes and (Inv)MixColumns steps. Because of this step-combination, as well as the high parallelism achieved by multiple units of random-access memory (RAM) and random-access/content-addressable memory (RA/CAM) arrays, IMCRYPTO achieves high throughput encryption and decryption without sacrificing area and power consumption. Additionally, due to the integration of a RISC-V core, IMCRYPTO offers programmability and flexibility. IMCRYPTO improves the throughput per area by a minimum (maximum) of 3.3x (223.1x) when compared to previous ASICs/IMC architectures for AES-128 encryption. Projections show added benefit from emerging technologies of up to 5.3x to the area-delay-power product of IMCRYPTO.

Zheyu Yan, Weiwen Jiang, Xiaobo Sharon Hu et al.

Differentiable neural architecture search (DNAS) is known for its capacity in the automatic generation of superior neural networks. However, DNAS based methods suffer from memory usage explosion when the search space expands, which may prevent them from running successfully on even advanced GPU platforms. On the other hand, reinforcement learning (RL) based methods, while being memory efficient, are extremely time-consuming. Combining the advantages of both types of methods, this paper presents RADARS, a scalable RL-aided DNAS framework that can explore large search spaces in a fast and memory-efficient manner. RADARS iteratively applies RL to prune undesired architecture candidates and identifies a promising subspace to carry out DNAS. Experiments using a workstation with 12 GB GPU memory show that on CIFAR-10 and ImageNet datasets, RADARS can achieve up to 3.41% higher accuracy with 2.5X search time reduction compared with a state-of-the-art RL-based method, while the two DNAS baselines cannot complete due to excessive memory usage or search time. To the best of the authors' knowledge, this is the first DNAS framework that can handle large search spaces with bounded memory usage.

Zheyu Yan, Da-Cheng Juan, Xiaobo Sharon Hu et al.

Emerging device-based Computing-in-memory (CiM) has been proved to be a promising candidate for high-energy efficiency deep neural network (DNN) computations. However, most emerging devices suffer uncertainty issues, resulting in a difference between actual data stored and the weight value it is designed to be. This leads to an accuracy drop from trained models to actually deployed platforms. In this work, we offer a thorough analysis of the effect of such uncertainties-induced changes in DNN models. To reduce the impact of device uncertainties, we propose UAE, an uncertainty-aware Neural Architecture Search scheme to identify a DNN model that is both accurate and robust against device uncertainties.

Dayane Reis, Jonathan Takeshita, Taeho Jung et al.

Homomorphic encryption (HE) allows direct computations on encrypted data. Despite numerous research efforts, the practicality of HE schemes remains to be demonstrated. In this regard, the enormous size of ciphertexts involved in HE computations degrades computational efficiency. Near-memory Processing (NMP) and Computing-in-memory (CiM) - paradigms where computation is done within the memory boundaries - represent architectural solutions for reducing latency and energy associated with data transfers in data-intensive applications such as HE. This paper introduces CiM-HE, a Computing-in-memory (CiM) architecture that can support operations for the B/FV scheme, a somewhat homomorphic encryption scheme for general computation. CiM-HE hardware consists of customized peripherals such as sense amplifiers, adders, bit-shifters, and sequencing circuits. The peripherals are based on CMOS technology, and could support computations with memory cells of different technologies. Circuit-level simulations are used to evaluate our CiM-HE framework assuming a 6T-SRAM memory. We compare our CiM-HE implementation against (i) two optimized CPU HE implementations, and (ii) an FPGA-based HE accelerator implementation. When compared to a CPU solution, CiM-HE obtains speedups between 4.6x and 9.1x, and energy savings between 266.4x and 532.8x for homomorphic multiplications (the most expensive HE operation). Also, a set of four end-to-end tasks, i.e., mean, variance, linear regression, and inference are up to 1.1x, 7.7x, 7.1x, and 7.5x faster (and 301.1x, 404.6x, 532.3x, and 532.8x more energy efficient). Compared to CPU-based HE in a previous work, CiM-HE obtain 14.3x speed-up and >2600x energy savings. Finally, our design offers 2.2x speed-up with 88.1x energy savings compared to a state-of-the-art FPGA-based accelerator.

Weiwen Jiang, Qiuwen Lou, Zheyu Yan et al.

Co-exploration of neural architectures and hardware design is promising to simultaneously optimize network accuracy and hardware efficiency. However, state-of-the-art neural architecture search algorithms for the co-exploration are dedicated for the conventional von-neumann computing architecture, whose performance is heavily limited by the well-known memory wall. In this paper, we are the first to bring the computing-in-memory architecture, which can easily transcend the memory wall, to interplay with the neural architecture search, aiming to find the most efficient neural architectures with high network accuracy and maximized hardware efficiency. Such a novel combination makes opportunities to boost performance, but also brings a bunch of challenges. The design space spans across multiple layers from device type, circuit topology to neural architecture. In addition, the performance may degrade in the presence of device variation. To address these challenges, we propose a cross-layer exploration framework, namely NACIM, which jointly explores device, circuit and architecture design space and takes device variation into consideration to find the most robust neural architectures. Experimental results demonstrate that NACIM can find the robust neural network with 0.45% accuracy loss in the presence of device variation, compared with a 76.44% loss from the state-of-the-art NAS without consideration of variation; in addition, NACIM achieves an energy efficiency up to 16.3 TOPs/W, 3.17X higher than the state-of-the-art NAS.

Xiaoming Chen, Jianxu Chen, Danny Z. Chen et al.

Convolution is a fundamental operation in many applications, such as computer vision, natural language processing, image processing, etc. Recent successes of convolutional neural networks in various deep learning applications put even higher demand on fast convolution. The high computation throughput and memory bandwidth of graphics processing units (GPUs) make GPUs a natural choice for accelerating convolution operations. However, maximally exploiting the available memory bandwidth of GPUs for convolution is a challenging task. This paper introduces a general model to address the mismatch between the memory bank width of GPUs and computation data width of threads. Based on this model, we develop two convolution kernels, one for the general case and the other for a special case with one input channel. By carefully optimizing memory access patterns and computation patterns, we design a communication-optimized kernel for the special case and a communication-reduced kernel for the general case. Experimental data based on implementations on Kepler GPUs show that our kernels achieve 5.16X and 35.5% average performance improvement over the latest cuDNN library, for the special case and the general case, respectively.