Chi-Ying Tsui

Liang Zhao, Kunming Shao, Zhipeng Liao et al.

FP8 low-precision formats have gained significant adoption in Transformer inference and training. However, existing digital compute-in-memory (DCIM) architectures face challenges in supporting variable FP8 aligned-mantissa bitwidths, as unified alignment strategies and fixed-precision multiply-accumulate (MAC) units struggle to handle input data with diverse distributions. This work presents a flexible FP8 DCIM accelerator with three innovations: (1) a dynamic shift-aware bitwidth prediction (DSBP) with on-the-fly input prediction that adaptively adjusts weight (2/4/6/8b) and input (2$\sim$12b) aligned-mantissa precision; (2) a FIFO-based input alignment unit (FIAU) replacing complex barrel shifters with pointer-based control; and (3) a precision-scalable INT MAC array achieving flexible weight precision with minimal overhead. Implemented in 28nm CMOS with a 64$\times$96 CIM array, the design achieves 20.4 TFLOPS/W for fixed E5M7, demonstrating 2.8$\times$ higher FP8 efficiency than previous work while supporting all FP8 formats. Results on Llama-7b show that the DSBP achieves higher efficiency than fixed bitwidth mode at the same accuracy level on both BoolQ and Winogrande datasets, with configurable parameters enabling flexible accuracy-efficiency trade-offs.

Linping Qu, Shenghui Song, Chi-Ying Tsui et al.

Because of its privacy-preserving capability, federated learning (FL) has attracted significant attention from both academia and industry. However, when being implemented over wireless networks, it is not clear how much communication error can be tolerated by FL. This paper investigates the robustness of FL to the uplink and downlink communication error. Our theoretical analysis reveals that the robustness depends on two critical parameters, namely the number of clients and the numerical range of model parameters. It is also shown that the uplink communication in FL can tolerate a higher bit error rate (BER) than downlink communication, and this difference is quantified by a proposed formula. The findings and theoretical analyses are further validated by extensive experiments.

Pingcheng Dong, Yonghao Tan, Xuejiao Liu et al.

This work presents a 55nm speculative decoding-based LLM accelerator with bumping-based face-to-face ReRAM-on-logic stacking technology. It features a local rotation unit for outlier-free low-bit quantization, a stacking-aware PNM architecture co-designed with blockwise vector quantization to reduce weight EMA overheads, and an adaptive parallel speculative decoding scheme with an out-of-order scheduler for high resource and bandwidth utilization. Our chip achieves 14.08-to-135.69token/s and 4.46-to-7.17x speedup over vanilla speculative decoding.

Linping Qu, Shenghui Song, Chi-Ying Tsui

In wireless federated learning (FL), the clients need to transmit the high-dimensional deep neural network (DNN) parameters through bandwidth-limited channels, which causes the communication latency issue. In this paper, we propose a layer-wise adaptive modulation scheme to save the communication latency. Unlike existing works which assign the same modulation level for all DNN layers, we consider the layers' importance which provides more freedom to save the latency. The proposed scheme can automatically decide the optimal modulation levels for different DNN layers. Experimental results show that the proposed scheme can save up to 73.9% of communication latency compared with the existing schemes.

Xiaoyu Gan, Jingbo Jiang, Jingyang Zhu et al.

Substantial efforts have been devoted to alleviating the impact of the long-tailed class distribution in federated learning. In this work, we observe an interesting phenomenon that certain weak classes consistently exist even for class-balanced learning. These weak classes, different from the minority classes in the previous works, are inherent to data and remain fairly consistent for various network structures, learning paradigms, and data partitioning methods. The inherent inter-class accuracy discrepancy can reach over 36.9% for federated learning on the FashionMNIST and CIFAR-10 datasets, even when the class distribution is balanced both globally and locally. In this study, we empirically analyze the potential reason for this phenomenon. Furthermore, a partial knowledge distillation (PKD) method is proposed to improve the model's classification accuracy for weak classes. In this approach, knowledge transfer is initiated upon the occurrence of specific misclassifications within certain weak classes. Experimental results show that the accuracy of weak classes can be improved by 10.7%, reducing the inherent inter-class discrepancy effectively.

Linping Qu, Yuyi Mao, Shenghui Song et al.

One of the most critical challenges for deploying distributed learning solutions, such as federated learning (FL), in wireless networks is the limited battery capacity of mobile clients. While it is a common belief that the major energy consumption of mobile clients comes from the uplink data transmission, this paper presents a novel finding, namely channel decoding also contributes significantly to the overall energy consumption of mobile clients in FL. Motivated by this new observation, we propose an energy-efficient adaptive channel decoding scheme that leverages the intrinsic robustness of FL to model errors. In particular, the robustness is exploited to reduce the energy consumption of channel decoders at mobile clients by adaptively adjusting the number of decoding iterations. We theoretically prove that wireless FL with communication errors can converge at the same rate as the case with error-free communication provided the bit error rate (BER) is properly constrained. An adaptive channel decoding scheme is then proposed to improve the energy efficiency of wireless FL systems. Experimental results demonstrate that the proposed method maintains the same learning accuracy while reducing the channel decoding energy consumption by ~20% when compared to an existing approach.

Linping Qu, Shenghui Song, Chi-Ying Tsui

Federated learning (FL) is a powerful machine learning paradigm which leverages the data as well as the computational resources of clients, while protecting clients' data privacy. However, the substantial model size and frequent aggregation between the server and clients result in significant communication overhead, making it challenging to deploy FL in resource-limited wireless networks. In this work, we aim to mitigate the communication overhead by using quantization. Previous research on quantization has primarily focused on the uplink communication, employing either fixed-bit quantization or adaptive quantization methods. In this work, we introduce a holistic approach by joint uplink and downlink adaptive quantization to reduce the communication overhead. In particular, we optimize the learning convergence by determining the optimal uplink and downlink quantization bit-length, with a communication energy constraint. Theoretical analysis shows that the optimal quantization levels depend on the range of model gradients or weights. Based on this insight, we propose a decreasing-trend quantization for the uplink and an increasing-trend quantization for the downlink, which aligns with the change of the model parameters during the training process. Experimental results show that, the proposed joint uplink and downlink adaptive quantization strategy can save up to 66.7% energy compared with the existing schemes.

Linping Qu, Shenghui Song, Chi-Ying Tsui

Federated learning (FL) is an emerging learning paradigm without violating users' privacy. However, large model size and frequent model aggregation cause serious communication bottleneck for FL. To reduce the communication volume, techniques such as model compression and quantization have been proposed. Besides the fixed-bit quantization, existing adaptive quantization schemes use ascending-trend quantization, where the quantization level increases with the training stages. In this paper, we first investigate the impact of quantization on model convergence, and show that the optimal quantization level is directly related to the range of the model updates. Given the model is supposed to converge with the progress of the training, the range of the model updates will gradually shrink, indicating that the quantization level should decrease with the training stages. Based on the theoretical analysis, a descending quantization scheme named FedDQ is proposed. Experimental results show that the proposed descending quantization scheme can save up to 65.2% of the communicated bit volume and up to 68% of the communication rounds, when compared with existing schemes.

Xizi Chen, Jingyang Zhu, Jingbo Jiang et al.

The unstructured sparsity after pruning poses a challenge to the efficient implementation of deep learning models in existing regular architectures like systolic arrays. On the other hand, coarse-grained structured pruning is suitable for implementation in regular architectures but tends to have higher accuracy loss than unstructured pruning when the pruned models are of the same size. In this work, we propose a model compression method based on a novel weight permutation scheme to fully exploit the fine-grained weight sparsity in the hardware design. Through permutation, the optimal arrangement of the weight matrix is obtained, and the sparse weight matrix is further compressed to a small and dense format to make full use of the hardware resources. Two pruning granularities are explored. In addition to the unstructured weight pruning, we also propose a more fine-grained subword-level pruning to further improve the compression performance. Compared to the state-of-the-art works, the matrix compression rate is significantly improved from 5.88x to 14.13x. As a result, the throughput and energy efficiency are improved by 2.75 and 1.86 times, respectively.

Jingbo Jiang, Xizi Chen, Chi-Ying Tsui

Recent literature has shown that convolutional neural networks (CNNs) with large kernels outperform vision transformers (ViTs) and CNNs with stacked small kernels in many computer vision tasks, such as object detection and image restoration. The Winograd transformation helps reduce the number of repetitive multiplications in convolution and is widely supported by many commercial AI processors. Researchers have proposed accelerating large kernel convolutions by linearly decomposing them into many small kernel convolutions and then sequentially accelerating each small kernel convolution with the Winograd algorithm. This work proposes a nested Winograd algorithm that iteratively decomposes a large kernel convolution into small kernel convolutions and proves it to be more effective than the linear decomposition Winograd transformation algorithm. Experiments show that compared to the linear decomposition Winograd algorithm, the proposed algorithm reduces the total number of multiplications by 1.4 to 10.5 times for computing 4x4 to 31x31 convolutions.

Jingyang Zhu, Jingbo Jiang, Xizi Chen et al.

Contemporary Deep Neural Network (DNN) contains millions of synaptic connections with tens to hundreds of layers. The large computation and memory requirements pose a challenge to the hardware design. In this work, we leverage the intrinsic activation sparsity of DNN to substantially reduce the execution cycles and the energy consumption. An end-to-end training algorithm is proposed to develop a lightweight run-time predictor for the output activation sparsity on the fly. From our experimental results, the computation overhead of the prediction phase can be reduced to less than 5% of the original feedforward phase with negligible accuracy loss. Furthermore, an energy-efficient hardware architecture, SparseNN, is proposed to exploit both the input and output sparsity. SparseNN is a scalable architecture with distributed memories and processing elements connected through a dedicated on-chip network. Compared with the state-of-the-art accelerators which only exploit the input sparsity, SparseNN can achieve a 10%-70% improvement in throughput and a power reduction of around 50%.