skip to main content


Title: AccPar: Tensor Partitioning for Heterogeneous Deep Learning Accelerators
Deep neural network (DNN) accelerators as an example of domain-specific architecture have demonstrated great success in DNN inference. However, the architecture acceleration for equally important DNN training has not yet been fully studied. With data forward, error backward and gradient calculation, DNN training is a more complicated process with higher computation and communication intensity. Because the recent research demonstrates a diminishing specialization return, namely, “accelerator wall”, we believe that a promising approach is to explore coarse-grained parallelism among multiple performance-bounded accelerators to support DNN training. Distributing computations on multiple heterogeneous accelerators to achieve high throughput and balanced execution, however, remaining challenging. We present ACCPAR, a principled and systematic method of determining the tensor partition among heterogeneous accelerator arrays. Compared to prior empirical or unsystematic methods, ACCPAR considers the complete tensor partition space and can reveal previously unknown new parallelism configurations. ACCPAR optimizes the performance based on a cost model that takes into account both computation and communication costs of a heterogeneous execution environment. Hence, our method can avoid the drawbacks of existing approaches that use communication as a proxy of the performance. The enhanced flexibility of tensor partitioning in ACCPAR allows the flexible ratio of computations to be distributed among accelerators with different performances. The proposed search algorithm is also applicable to the emerging multi-path patterns in modern DNNs such as ResNet. We simulate ACCPAR on a heterogeneous accelerator array composed of both TPU-v2 and TPU-v3 accelerators for the training of large-scale DNN models such as Alexnet, Vgg series and Resnet series. The average performance improvements of the state-of-the-art “one weird trick” (OWT) and HYPAR, and ACCPAR, normalized to the baseline data parallelism scheme where each accelerator replicates the model and processes different input data in parallel, are 2.98×, 3.78×, and 6.30×, respectively.  more » « less
Award ID(s):
1910299 1717885
NSF-PAR ID:
10179973
Author(s) / Creator(s):
; ; ; ; ;
Date Published:
Journal Name:
2020 IEEE International Symposium on High Performance Computer Architecture (HPCA)
Page Range / eLocation ID:
342 to 355
Format(s):
Medium: X
Sponsoring Org:
National Science Foundation
More Like this
  1. With the proliferation of low-cost sensors and the Internet of Things, the rate of producing data far exceeds the compute and storage capabilities of today’s infrastructure. Much of this data takes the form of time series, and in response, there has been increasing interest in the creation of time series archives in the last decade, along with the development and deployment of novel analysis methods to process the data. The general strategy has been to apply a plurality of similarity search mechanisms to various subsets and subsequences of time series data in order to identify repeated patterns and anomalies; however, the computational demands of these approaches renders them incompatible with today’s power-constrained embedded CPUs. To address this challenge, we present FA-LAMP, an FPGA-accelerated implementation of the Learned Approximate Matrix Profile (LAMP) algorithm, which predicts the correlation between streaming data sampled in real-time and a representative time series dataset used for training. FA-LAMP lends itself as a real-time solution for time series analysis problems such as classification. We present the implementation of FA-LAMP on both edge- and cloud-based prototypes. On the edge devices, FA-LAMP integrates accelerated computation as close as possible to IoT sensors, thereby eliminating the need to transmit and store data in the cloud for posterior analysis. On the cloud-based accelerators, FA-LAMP can execute multiple LAMP models on the same board, allowing simultaneous processing of incoming data from multiple data sources across a network. LAMP employs a Convolutional Neural Network (CNN) for prediction. This work investigates the challenges and limitations of deploying CNNs on FPGAs using the Xilinx Deep Learning Processor Unit (DPU) and the Vitis AI development environment. We expose several technical limitations of the DPU, while providing a mechanism to overcome them by attaching custom IP block accelerators to the architecture. We evaluate FA-LAMP using a low-cost Xilinx Ultra96-V2 FPGA as well as a cloud-based Xilinx Alveo U280 accelerator card and measure their performance against a prototypical LAMP deployment running on a Raspberry Pi 3, an Edge TPU, a GPU, a desktop CPU, and a server-class CPU. In the edge scenario, the Ultra96-V2 FPGA improved performance and energy consumption compared to the Raspberry Pi; in the cloud scenario, the server CPU and GPU outperformed the Alveo U280 accelerator card, while the desktop CPU achieved comparable performance; however, the Alveo card offered an order of magnitude lower energy consumption compared to the other four platforms. Our implementation is publicly available at https://github.com/aminiok1/lamp-alveo. 
    more » « less
  2. DNN training is extremely time-consuming, necessitating efficient multi-accelerator parallelization. Current approaches to parallelizing training primarily use intra-batch parallelization, where a single iteration of training is split over the available workers, but suffer from diminishing returns at higher worker counts. We present PipeDream, a system that adds inter-batch pipelining to intra-batch parallelism to further improve parallel training throughput, helping to better overlap computation with communication and reduce the amount of communication when possible. Unlike traditional pipelining, DNN training is bi-directional, where a forward pass through the computation graph is followed by a backward pass that uses state and intermediate data computed during the forward pass. Naïve pipelining can thus result in mismatches in state versions used in the forward and backward passes, or excessive pipeline flushes and lower hardware efficiency. To address these challenges, PipeDream versions model parameters for numerically correct gradient computations, and schedules forward and backward passes of different minibatches concurrently on different workers with minimal pipeline stalls. PipeDream also automatically partitions DNN layers among workers to balance work and minimize communication. Extensive experimentation with a range of DNN tasks, models, and hardware configurations shows that PipeDream trains models to high accuracy up to 5.3X faster than commonly used intra-batch parallelism techniques. 
    more » « less
  3. Dense matrix multiply (MM) serves as one of the most heavily used kernels in deep learning applications. To cope with the high computation demands of these applications, heterogeneous architectures featuring both FPGA and dedicated ASIC accelerators have emerged as promising platforms. For example, the AMD/Xilinx Versal ACAP architecture combines general-purpose CPU cores and programmable logic (PL) with AI Engine processors (AIE) optimized for AI/ML. An array of 400 AI Engine processors executing at 1 GHz can theoretically provide up to 6.4 TFLOPs performance for 32-bit floating-point (fp32) data. However, machine learning models often contain both large and small MM operations. While large MM operations can be parallelized efficiently across many cores, small MM operations typically cannot. In our investigation, we observe that executing some small MM layers from the BERT natural language processing model on a large, monolithic MM accelerator in Versal ACAP achieved less than 5% of the theoretical peak performance. Therefore, one key question arises: How can we design accelerators to fully use the abundant computation resources under limited communication bandwidth for end-to-end applications with multiple MM layers of diverse sizes? We identify the biggest system throughput bottleneck resulting from the mismatch of massive computation resources of one monolithic accelerator and the various MM layers of small sizes in the application. To resolve this problem, we propose the CHARM framework to compose multiple diverse MM accelerator architectures working concurrently towards different layers within one application. CHARM includes analytical models which guide design space exploration to determine accelerator partitions and layer scheduling. To facilitate the system designs, CHARM automatically generates code, enabling thorough onboard design verification. We deploy the CHARM framework for four different deep learning applications, including BERT, ViT, NCF, MLP, on the AMD/Xilinx Versal ACAP VCK190 evaluation board. Our experiments show that we achieve 1.46 TFLOPs, 1.61 TFLOPs, 1.74 TFLOPs, and 2.94 TFLOPs inference throughput for BERT, ViT, NCF, MLP, respectively, which obtain 5.40x, 32.51x, 1.00x and 1.00x throughput gains compared to one monolithic accelerator. 
    more » « less
  4. In pursuit of higher inference accuracy, deep neural network (DNN) models have significantly increased in complexity and size. To overcome the consequent computational challenges, scalable chiplet-based accelerators have been proposed. However, data communication using metallic-based interconnects in these chiplet-based DNN accelerators is becoming a primary obstacle to performance, energy efficiency, and scalability. The photonic interconnects can provide adequate data communication support due to some superior properties like low latency, high bandwidth and energy efficiency, and ease of broadcast communication. In this paper, we propose SPACX: a Silicon Photonics-based Chiplet ACcelerator for DNN inference applications. Specifically, SPACX includes a photonic network design that enables seamless single-chiplet and cross-chiplet broadcast communications, and a tailored dataflow that promotes data broadcast and maximizes parallelism. Furthermore, we explore the broadcast granularities of the photonic network and implications on system performance and energy efficiency. A flexible bandwidth allocation scheme is also proposed to dynamically adjust communication bandwidths for different types of data. Simulation results using several DNN models show that SPACX can achieve 78% and 75% reduction in execution time and energy, respectively, as compared to other state-of-the-art chiplet-based DNN accelerators. 
    more » « less
  5. Deconvolution is a key component in contemporary neural networks, especially generative adversarial networks (GANs) and fully convolutional networks (FCNs). Due to extra operations of deconvolution compared to convolution, considerable degradation of performance as well as energy efficiency is incurred when implementing deconvolution on the existing resistive random access memory (ReRAM)-based processing-in-memory (PIM) accelerators. In this work, we propose a ReRAM-based accelerator design, RED, for providing high-performance and low-energy deconvolution. We analyze the deconvolution execution on the existing ReRAM-based PIMs and utilize its interior computation pattern for design optimization. RED includes two major contributions: pixel-wise mapping scheme and zero-skipping data flow. Pixel-wise mapping scheme removes the zero insertion and performs convolutions over several ReRAM arrays and thus enables parallel computations with non-zero inputs. Zero-skipping data flow, assisted with customized input buffers design, enhances the computation parallelism and input data reuse. In evaluation, we compare RED against the existing ReRAM-based PIMs and CMOS-based counterpart with a variety of GAN and FCN models, each of which contains multiple deconvolution layers. The experimental results show that RED achieves a 4.0×-56.16× speedup and a 1.05×-18.17× energy efficiency improvement over previous related accelerator designs. 
    more » « less