skip to main content

Title: SPACX: Silicon Photonics-based Scalable Chiplet Accelerator for DNN Inference
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.
; ;
Award ID(s):
1702980 1901165
Publication Date:
Journal Name:
IEEE International Symposium on High-Performance Computer Architecture (HPCA)
Page Range or eLocation-ID:
831 to 845
Sponsoring Org:
National Science Foundation
More Like this
  1. Chiplet-based architectures have been proposed to scale computing systems for deep neural networks (DNNs). Prior work has shown that for the chiplet-based DNN accelerators, the electrical network connecting the chiplets poses a major challenge to system performance, energy consumption, and scalability. Some emerging interconnect technologies such as silicon photonics can potentially overcome the challenges facing electrical interconnects as photonic interconnects provide high bandwidth density, superior energy efficiency, and ease of implementing broadcast and multicast operations that are prevalent in DNN inference. In this paper, we propose a chiplet-based architecture named SPRINT for DNN inference. SPRINT uses a global buffer to simplify the data transmission between storage and computation, and includes two novel designs: (1) a reconfigurable photonic network that can support diverse communications in DNN inference with minimal implementation cost, and (2) a customized dataflow that exploits the ease of broadcast and multicast feature of photonic interconnects to support highly parallel DNN computations. Simulation studies using ResNet50 DNN model show that SPRINT achieves 46% and 61% execution time and energy consumption reduction, respectively, as compared to other state-of-the-art chiplet-based architectures with electrical or photonic interconnects.
  2. Deep learning has led to unprecedented successes in solving some very difficult problems in domains such as computer vision, natural language processing, and general pattern recognition. These achievements are the culmination of decades-long research into better training techniques and deeper neural network models, as well as improvements in hardware platforms that are used to train and execute the deep neural network models. Many application-specific integrated circuit (ASIC) hardware accelerators for deep learning have garnered interest in recent years due to their improved performance and energy-efficiency over conventional CPU and GPU architectures. However, these accelerators are constrained by fundamental bottlenecks due to (1) the slowdown in CMOS scaling, which has limited computational and performance-per-watt capabilities of emerging electronic processors; and (2) the use of metallic interconnects for data movement, which do not scale well and are a major cause of bandwidth, latency, and energy inefficiencies in almost every contemporary processor. Silicon photonics has emerged as a promising CMOS-compatible alternative to realize a new generation of deep learning accelerators that can use light for both communication and computation. This article surveys the landscape of silicon photonics to accelerate deep learning, with a coverage of developments across design abstractions in a bottom-up manner,more »to convey both the capabilities and limitations of the silicon photonics paradigm in the context of deep learning acceleration.« less
  3. With the ever growing complexity of high performance computing (HPC) systems to satisfy emerging application requirements (e.g., high memory bandwidth requirement for machine learning applications), the performance bottleneck in such systems has moved from being computation-centric to be more communication-centric. Silicon photonic interconnection networks have been proposed to address the aggressive communication requirements in HPC systems, to realize higher bandwidth, lower latency, and better energy efficiency. There have been many successful efforts on developing silicon photonic devices, integrated circuits, and architectures for HPC systems. Moreover, many efforts have been made to address and mitigate the impact of different challenges (e.g., fabrication process and thermal variations) in silicon photonic interconnects. However, most of these efforts have focused only on a single design layer in the system design space (e.g., device, circuit or architecture level). Therefore, there is often a gap between what a design technique can improve in one layer, and what it might impair in another one. In this paper, we discuss the promise of cross-layer design methodologies for HPC systems integrating silicon photonic interconnects. In particular, we discuss how such cross-layer design solutions based on cooperatively designing and exchanging design objectives among different system design layers can help achievemore »the best possible performance when integrating silicon photonics into HPC systems« less
  4. 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 amongmore »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.« less
  5. Domain specific neural network accelerators have garnered attention because of their improved energy efficiency and inference performance compared to CPUs and GPUs. Such accelerators are thus well suited for resource-constrained embedded systems. However, mapping sophisticated neural network models on these accelerators still entails significant energy and memory consumption, along with high inference time overhead. Binarized neural networks (BNNs), which utilize single-bit weights, represent an efficient way to implement and deploy neural network models on accelerators. In this paper, we present a novel optical-domain BNN accelerator, named ROBIN , which intelligently integrates heterogeneous microring resonator optical devices with complementary capabilities to efficiently implement the key functionalities in BNNs. We perform detailed fabrication-process variation analyses at the optical device level, explore efficient corrective tuning for these devices, and integrate circuit-level optimization to counter thermal variations. As a result, our proposed ROBIN architecture possesses the desirable traits of being robust, energy-efficient, low latency, and high throughput, when executing BNN models. Our analysis shows that ROBIN can outperform the best-known optical BNN accelerators and many electronic accelerators. Specifically, our energy-efficient ROBIN design exhibits energy-per-bit values that are ∼4 × lower than electronic BNN accelerators and ∼933 × lower than a recently proposed photonic BNNmore »accelerator, while a performance-efficient ROBIN design shows ∼3 × and ∼25 × better performance than electronic and photonic BNN accelerators, respectively.« less