Search

Benefiting from the self-attention mechanism, Transformer models have attained impressive contextual comprehension capabilities for lengthy texts. The requirements of high-throughput inference arise as the large language models (LLMs) become increasingly prevalent, which calls for large-scale token parallel processing (LTPP). However, existing dynamic sparse accelerators struggle to effectively handle LTPP, as they solely focus on separate stage optimization, and with most efforts confined to computational enhancements. By re-examining the end-to-end flow of dynamic sparse acceleration, we pinpoint an ever-overlooked opportunity that the LTPP can exploit the intrinsic coordination among stages to avoid excessive memory access and redundant computation. Motivated by our observation, we present SOFA, a cross-stage compute-memory efficient algorithm-hardware co-design, which is tailored to tackle the challenges posed by LTPP of Transformer inference effectively. We first propose a novel leading zero computing paradigm, which predicts attention sparsity by using log-based add-only operations to avoid the significant overhead of prediction. Then, a distributed sorting and a sorted updating FlashAttention mechanism are proposed with a cross-stage coordinated tiling principle, which enables fine-grained and lightweight coordination among stages, helping optimize memory access and latency. Further, we propose a SOFA accelerator to support these optimizations efficiently. Extensive experiments on 20 benchmarks show that SOFA achieves $9.5\times$ speed up and $71.5\times$ higher energy efficiency than Nvidia A100 GPU. Compared to 8 SOTA accelerators, SOFA achieves an average $15.8\times$ energy efficiency, $10.3\times$ area efficiency and $9.3\times$ speed up, respectively.

Deep learning (DL) models are piquing high interest and scaling at an unprecedented rate. To this end, a handful of tiled accelerators have been proposed to support such large-scale training tasks. However, these accelerators often incorporate numerous cores or tiles even extending to wafer-scale, substantial on-chip bandwidth, and distributed memory systems. This results in an exceedingly complex design space. Moreover, conducting actual training experiments to find optimal configurations is impractical due to time constraints. Hence, predicting the optimal mapping of various parallelisms to such tiled system architectures becomes crucial. In this study, leveraging an analysis of existing mainstream DL model training strategies, we introduce a performance simulator named PALM. PALM targets both the training and inference processes for tiled accelerators, aiming to inspire the design of current and future accelerators. Specifically, (i) we establish a scheduling mechanism among tiled accelerators based on an event-driven framework; (ii) we support user-configurable pipeline, tensor, and data parallelism on tiled accelerators, determining the absolute performance throughput under these parallelism strategies; (iii) we model the interaction of on-chip SRAM, NoC, and off-chip DRAM during operator execution. This work is available here: https://github.com/fangjh21/PALM., Comment: 11 pages

Given the increasing complexity of AI applications, traditional spatial architectures frequently fall short. Our analysis identifies a pattern of interconnected, multi-faceted tasks encompassing both AI and general computational processes. In response, we have conceptualized "Orchestrated AI Workflows," an approach that integrates various tasks with logic-driven decisions into dynamic, sophisticated workflows. Specifically, we find that the intrinsic Dual Dynamicity of Orchestrated AI Workflows, namely dynamic execution times and frequencies of Task Blocks, can be effectively represented using the Orchestrated Workflow Graph. Furthermore, the intrinsic Dual Dynamicity poses challenges to existing spatial architecture, namely Indiscriminate Resource Allocation, Reactive Load Rebalancing, and Contagious PEA Idleness. To overcome these challenges, we present Octopus, a scale-out spatial architecture and a suite of advanced scheduling strategies optimized for executing Orchestrated AI Workflows, such as the Discriminate Dual-Scheduling Mechanism, Adaptive TBU Scheduling Strategy, and Proactive Cluster Scheduling Strategy. Our evaluations demonstrate that Octopus significantly outperforms traditional architectures in handling the dynamic demands of Orchestrated AI Workflows, and possesses robust scalability in large scale hardware such as wafer-scale chip.

Nowadays, artificial intelligence (AI) technology with large models plays an increasingly important role in both academia and industry. It also brings a rapidly increasing demand for the computing power of the hardware. As the computing demand for AI continues to grow, the growth of hardware computing power has failed to keep up. This has become a significant factor restricting the development of AI. The augmentation of hardware computing power is mainly propelled by the escalation of transistor density and chip area. However, the former is impeded by the termination of the Moore's Law and Dennard scaling, and the latter is significantly restricted by the challenge of disrupting the legacy fabrication equipment and process. In recent years, advanced packaging technologies that have gradually matured are increasingly used to implement bigger chips that integrate multiple chiplets, while still providing interconnections with chip-level density and bandwidth. Compared to conventional high-performance computing paradigms such as multi-accelerator and datacenter-scale computing, Wafer-scale Computing shows remarkable advantages in communication bandwidth, integration density, and programmability potential. Not surprisingly, disruptive Wafer-scale Computing also brings unprecedented design challenges for hardware architecture, design-system-technology co-optimization, power and cooling systems, and compiler tool chain. At present, there are no comprehensive surveys summarizing the current state and design insights of Wafer-scale Computing. This paper aims to take the first step to help academia and industry review existing wafer-scale chips and essential technologies in a one-stop manner. So that people can conveniently grasp the basic knowledge and key points, understand the achievements and shortcomings of existing research, and contribute to this promising research direction.

With the cross-fertilization of applications and the ever-increasing scale of models, the efficiency and productivity of hardware computing architectures have become inadequate. This inadequacy further exacerbates issues in design flexibility, design complexity, development cycle, and development costs (4-d problems) in divergent scenarios. To address these challenges, this paper proposed a flexible design flow called DIAG based on plugin techniques. The proposed flow guides hardware development through four layers: definition(D), implementation(I), application(A), and generation(G). Furthermore, a versatile CGRA generator called WindMill is implemented, allowing for agile generation of customized hardware accelerators based on specific application demands. Applications and algorithm tasks from three aspects is experimented. In the case of reinforcement learning algorithm, a significant performance improvement of $2.3\times$ compared to GPU is achieved., Comment: 7 pages, 10 figures

Spatial architecture is a high-performance architecture that uses control flow graphs and data flow graphs as the computational model and producer/consumer models as the execution models. However, existing spatial architectures suffer from control flow handling challenges. Upon categorizing their PE execution models, we find that they lack autonomous, peer-to-peer, and temporally loosely-coupled control flow handling capability. This leads to limited performance in intensive control programs. A spatial architecture, Marionette, is proposed, with an explicit-designed control flow plane. The Control Flow Plane enables autonomous, peer-to-peer and temporally loosely-coupled control flow handling. The Proactive PE Configuration ensures timely and computation-overlapped configuration to improve handling Branch Divergence. The Agile PE Assignment enhance the pipeline performance of Imperfect Loops. We develop full stack of Marionette (ISA, compiler, simulator, RTL) and demonstrate that in a variety of challenging intensive control programs, compared to state-of-the-art spatial architectures, Marionette outperforms Softbrain, TIA, REVEL, and RipTide by geomean 2.88x, 3.38x, 1.55x, and 2.66x.

Diffusion Probability Models (DPMs) have made impressive advancements in various machine learning domains. However, achieving high-quality synthetic samples typically involves performing a large number of sampling steps, which impedes the possibility of real-time sample synthesis. Traditional accelerated sampling algorithms via knowledge distillation rely on pre-trained model weights and discrete time step scenarios, necessitating additional training sessions to achieve their goals. To address these issues, we propose the Catch-Up Distillation (CUD), which encourages the current moment output of the velocity estimation model ``catch up'' with its previous moment output. Specifically, CUD adjusts the original Ordinary Differential Equation (ODE) training objective to align the current moment output with both the ground truth label and the previous moment output, utilizing Runge-Kutta-based multi-step alignment distillation for precise ODE estimation while preventing asynchronous updates. Furthermore, we investigate the design space for CUDs under continuous time-step scenarios and analyze how to determine the suitable strategies. To demonstrate CUD's effectiveness, we conduct thorough ablation and comparison experiments on CIFAR-10, MNIST, and ImageNet-64. On CIFAR-10, we obtain a FID of 2.80 by sampling in 15 steps under one-session training and the new state-of-the-art FID of 3.37 by sampling in one step with additional training. This latter result necessitated only 620k iterations with a batch size of 128, in contrast to Consistency Distillation, which demanded 2100k iterations with a larger batch size of 256. Our code is released at https://anonymous.4open.science/r/Catch-Up-Distillation-E31F.

Deep learning applications are being transferred from the cloud to edge with the rapid development of embedded computing systems. In order to achieve higher energy efficiency with the limited resource budget, neural networks(NNs) must be carefully designed in two steps, the architecture design and the quantization policy choice. Neural Architecture Search(NAS) and Quantization have been proposed separately when deploying NNs onto embedded devices. However, taking the two steps individually is time-consuming and leads to a sub-optimal final deployment. To this end, we propose a novel neural network design framework called Hardware-aware Quantized Neural Architecture Search(HQNAS) framework which combines the NAS and Quantization together in a very efficient manner using weight-sharing and bit-sharing. It takes only 4 GPU hours to discover an outstanding NN policy on CIFAR10. It also takes only %10 GPU time to generate a comparable model on Imagenet compared to the traditional NAS method with 1.8x decrease of latency and a negligible accuracy loss of only 0.7%. Besides, our method can be adapted in a lifelong situation where the neural network needs to evolve occasionally due to changes of local data, environment and user preference.

Due to user privacy and regulatory restrictions, federate learning (FL) is proposed as a distributed learning framework for training deep neural networks (DNN) on decentralized data clients. Recent advancements in FL have applied Neural Architecture Search (NAS) to replace the predefined one-size-fit-all DNN model, which is not optimal for all tasks of various data distributions, with searchable DNN architectures. However, previous methods suffer from expensive communication cost rasied by frequent large model parameters transmission between the server and clients. Such difficulty is further amplified when combining NAS algorithms, which commonly require prohibitive computation and enormous model storage. Towards this end, we propose FAQS, an efficient personalized FL-NAS-Quantization framework to reduce the communication cost with three features: weight-sharing super kernels, bit-sharing quantization and masked transmission. FAQS has an affordable search time and demands very limited size of transmitted messages at each round. By setting different personlized pareto function loss on local clients, FAQS can yield heterogeneous hardware-aware models for various user preferences. Experimental results show that FAQS achieves average reduction of 1.58x in communication bandwith per round compared with normal FL framework and 4.51x compared with FL+NAS framwork.

Time Delay Neural Networks (TDNN)-based methods are widely used in dialect identification. However, in previous work with TDNN application, subtle variant is being neglected in different feature scales. To address this issue, we propose a new architecture, named dynamic multi-scale convolution, which consists of dynamic kernel convolution, local multi-scale learning, and global multi-scale pooling. Dynamic kernel convolution captures features between short-term and long-term context adaptively. Local multi-scale learning, which represents multi-scale features at a granular level, is able to increase the range of receptive fields for convolution operation. Besides, global multi-scale pooling is applied to aggregate features from different bottleneck layers in order to collect information from multiple aspects. The proposed architecture significantly outperforms state-of-the-art system on the AP20-OLR-dialect-task of oriental language recognition (OLR) challenge 2020, with the best average cost performance (Cavg) of 0.067 and the best equal error rate (EER) of 6.52%. Compared with the known best results, our method achieves 9% of Cavg and 45% of EER relative improvement, respectively. Furthermore, the parameters of proposed model are 91% fewer than the best known model.

Attention-based models have made tremendous progress on end-to-end automatic speech recognition(ASR) recently. However, the conventional transformer-based approaches usually generate the sequence results token by token from left to right, leaving the right-to-left contexts unexploited. In this work, we introduce a bidirectional speech transformer to utilize the different directional contexts simultaneously. Specifically, the outputs of our proposed transformer include a left-to-right target, and a right-to-left target. In inference stage, we use the introduced bidirectional beam search method, which can not only generate left-to-right candidates but also generate right-to-left candidates, and determine the best hypothesis by the score. To demonstrate our proposed speech transformer with a bidirectional decoder(STBD), we conduct extensive experiments on the AISHELL-1 dataset. The results of experiments show that STBD achieves a 3.6\% relative CER reduction(CERR) over the unidirectional speech transformer baseline. Besides, the strongest model in this paper called STBD-Big can achieve 6.64\% CER on the test set, without language model rescoring and any extra data augmentation strategies., Comment: Accepted by InterSpeech 2020

This paper describes the systems submitted by the department of electronic engineering, institute of microelectronics of Tsinghua university and TsingMicro Co. Ltd. (THUEE) to the NIST 2019 speaker recognition evaluation CTS challenge. Six subsystems, including etdnn/ams, ftdnn/as, eftdnn/ams, resnet, multitask and c-vector are developed in this evaluation., Comment: This is the system description of THUEE submitted to NIST SRE 2019

Despite the recent successes of deep neural networks, it remains challenging to achieve high precision keyword spotting task (KWS) on resource-constrained devices. In this study, we propose a novel context-aware and compact architecture for keyword spotting task. Based on residual connection and bottleneck structure, we design a compact and efficient network for KWS task. To leverage the long range dependencies and global context of the convolutional feature maps, the graph convolutional network is introduced to encode the non-local relations. By evaluated on the Google Speech Command Dataset, the proposed method achieves state-of-the-art performance and outperforms the prior works by a large margin with lower computational cost., Comment: Accepted by the IEEE Automatic Speech Recognition and Understanding Workshop(ASRU 2019)