Xuejiao Liu

Pingcheng Dong, Yonghao Tan, Dong Zhang et al.

Non-linear functions are prevalent in Transformers and their lightweight variants, incurring substantial and frequently underestimated hardware costs. Previous state-of-the-art works optimize these operations by piece-wise linear approximation and store the parameters in look-up tables (LUT), but most of them require unfriendly high-precision arithmetics such as FP/INT 32 and lack consideration of integer-only INT quantization. This paper proposed a genetic LUT-Approximation algorithm namely GQA-LUT that can automatically determine the parameters with quantization awareness. The results demonstrate that GQA-LUT achieves negligible degradation on the challenging semantic segmentation task for both vanilla and linear Transformer models. Besides, proposed GQA-LUT enables the employment of INT8-based LUT-Approximation that achieves an area savings of 81.3~81.7% and a power reduction of 79.3~80.2% compared to the high-precision FP/INT 32 alternatives. Code is available at https:// github.com/PingchengDong/GQA-LUT.

Yonghao Tan, Pingcheng Dong, Yongkun Wu et al.

DNN accelerators, significantly advanced by model compression and specialized dataflow techniques, have marked considerable progress. However, the frequent access of high-precision partial sums (PSUMs) leads to excessive memory demands in architectures utilizing input/weight stationary dataflows. Traditional compression strategies have typically overlooked PSUM quantization, which may account for 69% of power consumption. This study introduces a novel Additive Partial Sum Quantization (APSQ) method, seamlessly integrating PSUM accumulation into the quantization framework. A grouping strategy that combines APSQ with PSUM quantization enhanced by a reconfigurable architecture is further proposed. The APSQ performs nearly lossless on NLP and CV tasks across BERT, Segformer, and EfficientViT models while compressing PSUMs to INT8. This leads to a notable reduction in energy costs by 28-87%. Extended experiments on LLaMA2-7B demonstrate the potential of APSQ for large language models. Code is available at https://github.com/Yonghao-Tan/APSQ.

Songchen Ma, Hongyi Li, Weihao Zhang et al.

Mixture-of-Experts is a promising approach for edge AI with low-batch inference. Yet, on-device deployments often face limited on-chip memory and severe workload imbalance; the prevalent use of offloading further incurs off-chip memory access bottlenecks. Moreover, MoE sparsity and dynamic gating shift distributed strategies toward much finer granularity and introduce runtime scheduling considerations. Recently, high die-to-die bandwidth chiplet interconnects have created new opportunities for multi-chiplet systems to address workload imbalance and offloading bottlenecks with fine-grained scheduling. In this paper, we propose Fully Sharded Expert Data Parallelism, a parallelization paradigm specifically architected for low-batch MoE inference on multi-chiplet accelerators. FSE-DP attains adaptive computation-communication overlap and balanced load by orchestrating fine-grained, complementary expert streams along dynamic trajectories across high-bandwidth D2D links. The attendant dataflow complexity is tamed by a minimal, hardware-amenable set of virtualization rules and a lightweight scheduling algorithm. Our approach achieves 1.22 to 2.00 times speedup over state-of-the-art baselines and saves up to 78.8 percent on-chip memory.

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.

Yifan Qin, Zhenge Jia, Zheyu Yan et al.

This paper proposes an ultra-low power, mixed-bit-width sparse convolutional neural network (CNN) accelerator to accelerate ventricular arrhythmia (VA) detection. The chip achieves 50% sparsity in a quantized 1D CNN using a sparse processing element (SPE) architecture. Measurement on the prototype chip TSMC 40nm CMOS low-power (LP) process for the VA classification task demonstrates that it consumes 10.60 $μ$W of power while achieving a performance of 150 GOPS and a diagnostic accuracy of 99.95%. The computation power density is only 0.57 $μ$W/mm$^2$, which is 14.23X smaller than state-of-the-art works, making it highly suitable for implantable and wearable medical devices.

Meiyu Huang, Yao Xu, Lixin Qian et al.

Deep learning techniques have made an increasing impact on the field of remote sensing. However, deep neural networks based fusion of multimodal data from different remote sensors with heterogenous characteristics has not been fully explored, due to the lack of availability of big amounts of perfectly aligned multi-sensor image data with diverse scenes of high resolutions, especially for synthetic aperture radar (SAR) data and optical imagery. To promote the development of deep learning based SAR-optical fusion approaches, we release the QXS-SAROPT dataset, which contains 20,000 pairs of SAR-optical image patches. We obtain the SAR patches from SAR satellite GaoFen-3 images and the optical patches from Google Earth images. These images cover three port cities: San Diego, Shanghai and Qingdao. Here, we present a detailed introduction of the construction of the dataset, and show its two representative exemplary applications, namely SAR-optical image matching and SAR ship detection boosted by cross-modal information from optical images. As a large open SAR-optical dataset with multiple scenes of a high resolution, we believe QXS-SAROPT will be of potential value for further research in SAR-optical data fusion technology based on deep learning.

Wei Bao, Meiyu Huang, Yaqin Zhang et al.

Deep learning methods have made significant progress in ship detection in synthetic aperture radar (SAR) images. The pretraining technique is usually adopted to support deep neural networks-based SAR ship detectors due to the scarce labeled SAR images. However, directly leveraging ImageNet pretraining is hardly to obtain a good ship detector because of different imaging perspective and geometry. In this paper, to resolve the problem of inconsistent imaging perspective between ImageNet and earth observations, we propose an optical ship detector (OSD) pretraining technique, which transfers the characteristics of ships in earth observations to SAR images from a large-scale aerial image dataset. On the other hand, to handle the problem of different imaging geometry between optical and SAR images, we propose an optical-SAR matching (OSM) pretraining technique, which transfers plentiful texture features from optical images to SAR images by common representation learning on the optical-SAR matching task. Finally, observing that the OSD pretraining based SAR ship detector has a better recall on sea area while the OSM pretraining based SAR ship detector can reduce false alarms on land area, we combine the predictions of the two detectors through weighted boxes fusion to further improve detection results. Extensive experiments on four SAR ship detection datasets and two representative CNN-based detection benchmarks are conducted to show the effectiveness and complementarity of the two proposed detectors, and the state-of-the-art performance of the combination of the two detectors. The proposed method won the sixth place of ship detection in SAR images in 2020 Gaofen challenge.

Xuejiao Liu, Xueshuang Xiang

Generative adversarial networks (GANs) have been widely used and have achieved competitive results in semi-supervised learning. This paper theoretically analyzes how GAN-based semi-supervised learning (GAN-SSL) works. We first prove that, given a fixed generator, optimizing the discriminator of GAN-SSL is equivalent to optimizing that of supervised learning. Thus, the optimal discriminator in GAN-SSL is expected to be perfect on labeled data. Then, if the perfect discriminator can further cause the optimization objective to reach its theoretical maximum, the optimal generator will match the true data distribution. Since it is impossible to reach the theoretical maximum in practice, one cannot expect to obtain a perfect generator for generating data, which is apparently different from the objective of GANs. Furthermore, if the labeled data can traverse all connected subdomains of the data manifold, which is reasonable in semi-supervised classification, we additionally expect the optimal discriminator in GAN-SSL to also be perfect on unlabeled data. In conclusion, the minimax optimization in GAN-SSL will theoretically output a perfect discriminator on both labeled and unlabeled data by unexpectedly learning an imperfect generator, i.e., GAN-SSL can effectively improve the generalization ability of the discriminator by leveraging unlabeled information.

Xuejiao Liu, Yao Xu, Xueshuang Xiang

Generative adversarial networks (GANs) have attracted intense interest in the field of generative models. However, few investigations focusing either on the theoretical analysis or on algorithm design for the approximation ability of the generator of GANs have been reported. This paper will first theoretically analyze GANs' approximation property. Similar to the universal approximation property of the fully connected neural networks with one hidden layer, we prove that the generator with the input latent variable in GANs can universally approximate the potential data distribution given the increasing hidden neurons. Furthermore, we propose an approach named stochastic data generation (SDG) to enhance GANs'approximation ability. Our approach is based on the simple idea of imposing randomness through data generation in GANs by a prior distribution on the conditional probability between the layers. SDG approach can be easily implemented by using the reparameterization trick. The experimental results on synthetic dataset verify the improved approximation ability obtained by this SDG approach. In the practical dataset, four GANs using SDG can also outperform the corresponding traditional GANs when the model architectures are smaller.