This paper addresses design of accelerators using systolic architectures for training of neural networks using a novel gradient interleaving approach. Training the neural network involves backpropagation of error and computation of gradients with respect to the activation functions and weights. It is shown that the gradient with respect to the activation function can be computed using a weight-stationary systolic array while the gradient with respect to the weights can be computed using an output-stationary systolic array. The novelty of the proposed approach lies in interleaving the computations of these two gradients to the same configurable systolic array. This results in reuse of the variables from one computation to the other and eliminates unnecessary memory accesses. The proposed approach leads to 1.4–2.2× savings in terms of number of cycles and 1.9× savings in terms of memory accesses. Thus, the proposed accelerator reduces latency and energy consumption.
more »
« less
Polymorphic Accelerators for Deep Neural Networks
Deep neural networks (DNNs) come with many forms, such as convolutional neural networks, multilayer perceptron and recurrent neural networks, to meet diverse needs of machine learning applications. However, existing DNN accelerator designs, when used to execute multiple neural networks, suffer from underutilization of processing elements, heavy feature map traffic, and large area overhead. In this paper, we propose a novel approach, Polymorphic Accelerators, to address the flexibility issue fundamentally. We introduce the abstraction of logical accelerators to decouple the fixed mapping with physical resources. Three procedures are proposed that work collaboratively to reconfigure the accelerator for the current network that is being executed and to enable cross-layer data reuse among logical accelerators. Evaluation results show that the proposed approach achieves significant improvement in data reuse, inference latency and performance, e.g., 1.52x and 1.63x increase in throughput compared with state-of-the-art flexible dataflow approach and resource partitioning approach, respectively. This demonstrates the effectiveness and promise of polymorphic accelerator architecture.
more »
« less
- Award ID(s):
- 1750047
- NSF-PAR ID:
- 10315545
- Date Published:
- Journal Name:
- IEEE Transactions on Computers
- ISSN:
- 0018-9340
- Format(s):
- Medium: X
- Sponsoring Org:
- National Science Foundation
More Like this
-
-
Spiking Neural Networks (SNNs) are brain- inspired computing models incorporating unique temporal dynamics and event-driven processing. Rich dynamics in both space and time offer great challenges and opportunities for efficient processing of sparse spatiotemporal data compared with conventional artificial neural networks (ANNs). Specifically, the additional overheads for handling the added temporal dimension limit the computational capabilities of neuromorphic accelerators. Iterative processing at every time-point with sparse inputs in a temporally sequential manner not only degrades the utilization of the systolic array but also intensifies data movement.In this work, we propose a novel technique and architecture that significantly improve utilization and data movement while efficiently handling temporal sparsity of SNNs on systolic arrays. Unlike time-sequential processing in conventional SNN accelerators, we pack multiple time points into a single time window (TW) and process the computations induced by active synaptic inputs falling under several TWs in parallel, leading to the proposed parallel time batching. It allows weight reuse across multiple time points and enhances the utilization of the systolic array with reduced idling of processing elements, overcoming the irregularity of sparse firing activities. We optimize the granularity of time-domain processing, i.e., the TW size, which significantly impacts the data reuse and utilization. We further boost the utilization efficiency by simultaneously scheduling non-overlapping sparse spiking activities onto the array. The proposed architectures offer a unifying solution for general spiking neural networks with commonly exhibited temporal sparsity, a key challenge in hardware acceleration, delivering 248X energy-delay product (EDP) improvement on average compared to an SNN baseline for accelerating various networks. Compared to ANN based accelerators, our approach improves EDP by 47X on the CIFAR10 dataset.more » « less
-
This paper addresses the design of accelerators using systolic architectures to train convolutional neural networks using a novel gradient interleaving approach. Training the neural network involves computation and backpropagation of gradients of error with respect to the activation functions and weights. It is shown that the gradient with respect to the activation function can be computed using a weight-stationary systolic array, while the gradient with respect to the weights can be computed using an output-stationary systolic array. The novelty of the proposed approach lies in interleaving the computations of these two gradients on the same configurable systolic array. This results in the reuse of the variables from one computation to the other and eliminates unnecessary memory accesses and energy consumption associated with these memory accesses. The proposed approach leads to 1.4−2.2× savings in terms of the number of cycles and 1.9× savings in terms of memory accesses in the fully-connected layer. Furthermore, the proposed method uses up to 25% fewer cycles and memory accesses, and 16% less energy than baseline implementations for state-of-the-art CNNs. Under iso-area comparisons, for Inception-v4, compared to weight-stationary (WS), Intergrad achieves 12% savings in energy, 17% savings in memory, and 4% savings in cycles. Savings for Densenet-264 are 18% , 26% , and 27% with respect to energy, memory, and cycles, respectively. Thus, the proposed novel accelerator architecture reduces the latency and energy consumption for training deep neural networks.more » « less
-
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 BNN accelerator, while a performance-efficient ROBIN design shows ∼3 × and ∼25 × better performance than electronic and photonic BNN accelerators, respectively.more » « less
-
MAESTRO: A Data-Centric Approach to Understand Reuse, Performance, and Hardware Cost of DNN MappingsThe efficiency of an accelerator depends on three factors—mapping, deep neural network (DNN) layers, and hardware—constructing extremely complicated design space of DNN accelerators. To demystify such complicated design space and guide the DNN accelerator design for better efficiency, we propose an analytical cost model, MAESTRO. MAESTRO receives DNN model description and hardware resources information as a list, and mapping described in a data-centric representation we propose as inputs. The data centric representation consists of three directives that enable concise description of mappings in a compiler-friendly form. MAESTRO analyzes various forms of data reuse in an accelerator based on inputs quickly and generates more than 20 statistics including total latency, energy, throughput, etc., as outputs. MAESTRO’s fast analysis enables various optimization tools for DNN accelerators such as hardware design exploration tool we present as an example.more » « less
- Free Publicly Accessible Full Text
-
Accepted Manuscript1.0
- Journal Article:
- https://doi.org/10.1109/TC.2020.3048624
